Parameter-Efficient Fine-Tuning: A New Paradigm for Advancing Medical Image Analysis

1. Introduction: The Imperative for Efficiency in Adapting Foundational Models for Medical Imaging

The advent of foundation models, pre-trained on extensive and diverse datasets, has marked a significant turning point in artificial intelligence, with profound implications for medical image analysis.¹ These models, capable of learning highly generalizable representations, offer unprecedented potential for tackling complex diagnostic, prognostic, and analytical tasks within the medical domain. The dominant “pre-train fine-tune” paradigm allows these powerful generalist models to be adapted for specialized medical applications, promising to enhance clinical workflows and patient outcomes.¹

However, the widespread adoption of these large-scale models in medicine is significantly hampered by the challenges associated with full fine-tuning (FT). This traditional adaptation process, which involves updating all model parameters, is computationally intensive, demanding substantial graphical processing unit (GPU) resources, extensive training times, and considerable energy consumption.³ Such resource demands create a formidable barrier, particularly for many medical research institutions and clinical settings that may not possess the requisite high-performance computing infrastructure.

Furthermore, the medical imaging domain is often characterized by data scarcity. The curation of large, high-quality, annotated medical datasets is a perennial challenge due to stringent patient privacy regulations (e.g., HIPAA, GDPR), the high cost and time involved in expert annotation by clinicians, and the inherent rarity of certain diseases or conditions.¹ Attempting to fully fine-tune massive foundation models on such limited datasets significantly increases the risk of overfitting, where the model learns to perform well on the training data but fails to generalize to new, unseen patient cases. Beyond computational and data limitations, the practicalities of model management also pose a hurdle. Storing and maintaining separate, full-sized copies of multi-billion parameter models for each specific medical task, imaging modality, or sub-specialty is inefficient, costly, and logistically complex.⁵

In this context, Parameter-Efficient Fine-Tuning (PEFT) techniques have emerged as a strategic and transformative solution. PEFT encompasses a collection of methods designed to adapt large pre-trained models by fine-tuning only a small fraction of their parameters, or by introducing a minimal number of new, trainable parameters, while keeping the vast majority of the original model weights frozen.³ This approach drastically reduces the computational burden, memory footprint, and storage overhead associated with model adaptation, making the power of foundation models more accessible and sustainable, especially within the resource-aware medical field.¹ Crucially, PEFT aims to achieve this efficiency while maintaining, and in some cases even surpassing, the performance of full fine-tuning, particularly in data-constrained environments.¹

2. Understanding Parameter-Efficient Fine-Tuning (PEFT) Techniques

Parameter-Efficient Fine-Tuning (PEFT) refers to a suite of transfer learning methodologies specifically engineered to adapt large pre-trained models (LPMs), such as foundation models, to new downstream tasks or datasets with minimal computational and storage overhead.³ The core principle of PEFT is to freeze the vast majority of the pre-trained model’s parameters and update only a small, strategically chosen subset, or introduce a limited number of new, trainable parameters.¹² The primary objectives are manifold: to drastically reduce the number of trainable parameters compared to full fine-tuning, thereby lowering memory and storage requirements; to decrease the computational resources (GPU time, energy) needed for adaptation; and to achieve this efficiency while maintaining or even improving task-specific performance and generalization capabilities.³

2.1. Taxonomy of PEFT Methods

The landscape of PEFT is diverse and rapidly evolving, with methods generally classifiable into several broad categories based on how they modify or interact with the pre-trained model.³ A common taxonomy includes:

Additive PEFT: These techniques involve augmenting the pre-trained model architecture by injecting new, trainable modules or parameters. The original weights of the pre-trained model remain frozen, and only these newly added components are optimized during fine-tuning.³ This category includes:

• Adapters: Small neural network modules inserted between the layers of a pre-trained model.³
• Soft Prompts: Learnable continuous vector sequences prepended or interspersed within the input or intermediate representations to guide the model’s behavior. This includes methods like Prefix-Tuning, Prompt-Tuning, and P-Tuning.³

Other Additive Methods: Techniques like (IA)³ (Infused Adapter by Inhibiting and Amplifying Inner Activations) and SSF (Scaling and Shifting Features) modify activations using learned scaling factors and biases.³

• Selective PEFT: This approach fine-tunes only a very small subset of the LPM’s existing parameters. The selection of which parameters to tune can be based on heuristics, prior knowledge, or learned masks.³
  Examples include BitFit, which fine-tunes only the bias terms of the network, and methods employing structured or unstructured pruning masks to identify critical parameters for updates (e.g., FishMask, Diff Pruning).³
• Reparameterized PEFT: These methods involve transforming the model’s parameters, typically through low-rank decomposition, to make training more efficient. Only the parameters of the low-dimensional reparameterization are tuned. For inference, these can often be merged back into the original model structure to avoid latency.³ The most prominent example is LoRA (Low-Rank Adaptation) and its numerous variants.³

• Hybrid PEFT: As the name suggests, these methods combine elements from two or more of the above categories to leverage their respective strengths and potentially achieve superior performance or efficiency.³
• Quantization PEFT: This approach integrates model quantization techniques with PEFT. Quantization reduces the precision of model weights and activations (e.g., from 32-bit floating point to 8-bit integer or 4-bit), thereby decreasing memory footprint and potentially speeding up computation. It is often combined with other PEFTs, a notable example being QLoRA.³
• Multi-task PEFT: These methods are specifically designed for multi-task learning scenarios, often employing shared backbones with task-specific PEFT modules (like dynamic adapters) to enable efficient learning across multiple related tasks.³

2.2. In-depth Elaboration on Key PEFT Strategies

Several PEFT strategies have gained particular prominence due to their effectiveness and versatility.

2.2.1. Low-Rank Adaptation (LoRA) and its Variants

Low-Rank Adaptation (LoRA) has become one of the most popular PEFT techniques, especially for large language models and, increasingly, for vision models.⁸

• Mechanism: LoRA operates on the principle that the change in weights during model adaptation (ΔW) has a low intrinsic rank. Instead of fine-tuning the full weight matrix W of a pre-trained layer, LoRA freezes W and introduces two smaller, trainable “rank decomposition” matrices, A∈Rd×r and B∈Rr×k, where r is the rank and is typically much smaller than the original dimensions d and k (i.e., r≪min(d,k)). The update to the layer’s output h is then given by h=xW+x(BA), where x is the input to the layer.¹⁹ Matrix A is usually initialized with a random Gaussian distribution, while matrix B is initialized to zero. This ensures that at the beginning of training, the PEFT module BA is zero, and the model’s behavior is identical to the pre-trained model. Only A and B are updated during fine-tuning.
• Advantages: A key advantage of LoRA is that the trained BA matrices can be merged with the original weight matrix W (i.e., Wadapted​=W+BA) after training, resulting in no additional inference latency compared to the original model. It also dramatically reduces the number of trainable parameters.
• Variants:
  • DoRA (Weight-Decomposed Low-Rank Adaptation): DoRA aims to bridge the performance gap sometimes observed between LoRA and full fine-tuning by addressing a key difference in how they update weights.²⁴ It decomposes the pre-trained weight matrix W into its magnitude (a scalar m) and direction (V). Fine-tuning then involves learning updates for both components, where LoRA is specifically employed for updating the directional component V. This approach is designed to enhance the learning capacity and training stability of LoRA without adding inference overhead.²⁴
  • QLoRA (Quantized Low-Rank Adaptation): QLoRA pushes the boundaries of memory efficiency further by combining LoRA with quantization.³ In QLoRA, the large pre-trained model weights (W) are quantized to a very low precision (e.g., 4-bit NormalFloat), significantly reducing their memory footprint. The LoRA adapters (A and B) are then fine-tuned in a higher precision (e.g., 16-bit BrainFloat). Techniques like double quantization (quantizing the quantization constants themselves) and paged optimizers (to manage memory spikes during gradient checkpointing) are used to maintain performance and stability despite the aggressive quantization.²²

2.2.2. Adapter Tuning

Adapter tuning was one of the early and influential PEFT methods, demonstrating strong performance with high parameter efficiency.⁵

• Mechanism: Adapter tuning involves inserting small, task-specific neural network layers, known as “adapter modules,” within each layer (or selected layers) of a pre-trained network, such as a Transformer.¹¹ The parameters of the original pre-trained model are kept frozen. These adapter modules typically feature a bottleneck architecture: they first project the input features from the preceding layer down to a smaller dimension (the bottleneck dimension, m), apply a non-linear activation function, and then project the features back up to the original dimension (d).¹¹ Adapters are initialized to be near-identity functions at the start of training, ensuring that the pre-trained model’s initial behavior is preserved. Only the parameters of these adapter modules are trained for each downstream task.
• Advantages: Adapter tuning offers high parameter efficiency, as the number of parameters in each adapter module (2md+d+m) can be very small if m≪d. It promotes modularity, as different tasks can have their own adapters, which can be “plugged in” or “swapped out” as needed, allowing a single pre-trained backbone to serve multiple tasks. They generally yield good performance, often close to full fine-tuning.¹¹

2.2.3. Prompt-Based Tuning

Prompt-based tuning methods adapt LPMs by manipulating their input or internal representations using learnable “prompts,” rather than modifying the model’s architecture or core weights.⁴

• Prompt Tuning (PT): This technique prepends a sequence of continuous, learnable vectors, often called “soft prompts” or “virtual tokens,” to the input token embeddings of the LPM.¹⁶ These soft prompts are optimized end-to-end for a specific downstream task, while the parameters of the LPM itself remain frozen. The length of the soft prompt is a hyperparameter.
• Prefix Tuning: Prefix tuning is similar in spirit to prompt tuning but extends the concept by prepending learnable continuous vectors (prefixes) not just at the input layer but to the keys and values of the attention mechanism in every layer of a Transformer model.¹⁸ An MLP is often used to parameterize these prefix vectors to improve training stability and expressiveness. The original LPM parameters are frozen.
• Advantages: Both prompt tuning and prefix tuning are extremely parameter-efficient, as they only require optimizing the small set of prompt/prefix parameters. They have shown particular strength in natural language generation tasks and can effectively steer the behavior of very large models.

The proliferation of these PEFT methods, each with unique mechanisms and characteristics, underscores a critical consideration: there is no universally superior PEFT technique. The choice of an appropriate method is contingent upon a nuanced understanding of the specific application’s constraints and objectives. Factors such as the available computational memory, permissible inference latency, the architecture of the foundation model being adapted, and the inherent nature of the downstream medical task all play pivotal roles in this decision.³ For instance, LoRA’s advantage of zero added inference latency makes it attractive for real-time applications, whereas adapter modules might introduce slight delays due to the extra layers. QLoRA offers substantial memory savings, potentially enabling the fine-tuning of larger models on less powerful hardware, but this might come with a marginal trade-off in performance compared to methods using higher precision. Consequently, practitioners must meticulously evaluate these trade-offs. For deployment on resource-limited edge devices common in medical settings (e.g., portable ultrasound systems or wearable sensors), QLoRA or LoRA might be preferred, even if a more complex adapter configuration could theoretically yield slightly higher accuracy in a less constrained environment. This practical reality suggests that future research should not only focus on inventing novel PEFT algorithms but also on establishing comprehensive benchmarks and clear, evidence-based guidelines to aid in the selection of PEFT strategies tailored to diverse medical imaging scenarios. The development of automated systems, sometimes referred to as “AutoPEFT” ³, which can intelligently select and configure PEFT methods based on task and resource profiles, represents a promising avenue to navigate this complexity.

A notable trend in the evolution of PEFT is the increasing convergence with model compression techniques. While PEFT primarily aims to reduce the number of trainable parameters for efficient adaptation ³, model compression (e.g., quantization, pruning) targets the reduction of the overall model size and computational demands of the entire model, including its frozen parts.³ Techniques like QLoRA explicitly bridge these two domains by applying aggressive quantization to the frozen base model weights while fine-tuning LoRA adapters.³ This synergy is crucial because efficient adaptation increasingly involves not just deciding which parameters to tune, but also optimizing how the entire model—both its static and dynamic components—is represented, stored, and processed. This holistic approach to efficiency is paramount for deploying sophisticated foundation models on the often resource-constrained hardware found in clinical environments, such as portable diagnostic tools or embedded systems in medical devices. It is anticipated that future PEFT methodologies will more deeply integrate advanced compression techniques as a standard feature, rather than an optional add-on.

Furthermore, the trajectory of PEFT development indicates a shift from general-purpose techniques to more domain- and task-aware strategies. Early PEFT methods like standard Adapters or LoRA were conceived as broadly applicable across various tasks and model architectures.¹¹ However, more recent research explores PEFT variants that are either specifically designed for, or demonstrate particular aptitude in, certain contexts. Examples include visual prompt tuning methods tailored for computer vision tasks ³⁰, or techniques like DoRA that consider the structural properties of weight matrices within neural networks.²⁴ As foundation models grow in complexity and are applied to highly specialized and diverse tasks, such as those encountered in medical imaging, there is an escalating need for PEFT techniques that are not only parameter-efficient but also semantically aware of the specific characteristics of the task or data. For instance, an optimal PEFT strategy for segmenting intricate, fine-grained anatomical structures in 3D computed tomography (CT) scans might differ considerably from one best suited for classifying diffuse pathologies in 2D X-ray images. This points towards a future where PEFT methods might be co-designed with model architectures or specific medical applications in mind, potentially incorporating mechanisms to leverage known medical priors, imaging physics, or the unique statistical properties of medical data, rather than being applied as a generic, post-hoc efficiency layer.

To provide a clearer comparative perspective, Table 1 summarizes the key characteristics of major PEFT techniques.

Table 1: Comparative Overview of Major PEFT Techniques

3. PEFT in Action: Transforming Medical Image Analysis

The theoretical advantages of PEFT are increasingly being validated through practical applications across a wide spectrum of medical imaging tasks and modalities. These techniques are enabling researchers and clinicians to harness the power of large foundation models for nuanced medical analysis, which was previously hindered by computational and data limitations.

3.1. Applications in Medical Image Classification

PEFT methods have demonstrated considerable utility in this area. For instance, studies have benchmarked a variety of PEFT algorithms (up to 17 distinct methods) on diverse medical classification tasks, utilizing datasets of varying sizes, modalities (including X-ray, CT, MRI), and complexities.¹ These evaluations consistently highlight the effectiveness of PEFT, particularly in low-data regimes where performance gains of up to 22% over baseline fine-tuning approaches have been reported.¹ This is crucial for medical imaging, where large labeled datasets are often a luxury.

Specific PEFT techniques are being tailored for medical classification. Dynamic Visual Prompt Tuning (DVPT) has been applied to tasks like diabetic retinopathy grading from 2D retinal images, showing strong performance with minimal trainable parameters.³⁰ Another novel approach, Embedded Prompt Tuning (EPT), has been proposed to address limitations in how prompts are introduced in Transformer architectures. EPT has demonstrated superior performance compared to several state-of-the-art fine-tuning methods in few-shot medical image classification scenarios, completing the fine-tuning process efficiently.³⁵ These examples underscore PEFT’s role in making advanced classification models more adaptable and effective for specific medical diagnostic challenges.

3.2. Advancements in Medical Image Segmentation using PEFT

Foundation models like the Segment Anything Model (SAM) are being fine-tuned for biomedical image segmentation (including microscopy and various medical imaging modalities) using PEFT techniques such as LoRA and QLoRA.²⁰ A key observation from such studies is that the strategic placement of PEFT layers within the model architecture can be more critical for achieving efficiency and performance than the specific type of PEFT layer used.²⁰ This highlights the need for careful architectural consideration when applying PEFT to complex vision models.

For multi-modal segmentation, the PEMMA (Parameter-Efficient Multi-Modal Adaptation) framework utilizes LoRA or its variant DoRA to adapt transformer-based segmentation models, initially trained on CT scans, for use with PET scans and even for prognostic tasks using additional data like electronic health records (EHR).³² Applied to datasets like HECKTOR for head and neck cancer, PEMMA has achieved performance comparable to more data-intensive early fusion methods while using only about 8% of the trainable parameters, and has shown significant improvements in Dice scores for PET scan segmentation.³²

DVPT has also been extended to segmentation tasks, including polyp segmentation from colonoscopy images, skin lesion segmentation from dermoscopic images, cardiac chamber segmentation from MRI, and multi-organ segmentation from CT scans.³⁰ Furthermore, the PRS-Med framework employs LoRA to fine-tune the LLaVA-Med vision-language model in conjunction with a TinySAM image encoder for tumor segmentation and, notably, position reasoning across a wide array of modalities including CT, MRI, X-ray, ultrasound, and endoscopy.³⁴ Even LLMs fine-tuned with QLoRA for decision support can indirectly aid segmentation tasks by providing contextual understanding or identifying regions of interest from textual descriptions or patient history.²¹

3.3. PEFT for Medical Image Generation, Synthesis, and Reconstruction

Beyond discriminative tasks, PEFT is also making inroads into generative modeling for medical imaging, which includes image synthesis (creating realistic medical images from scratch or based on conditions like text prompts) and image reconstruction (improving the quality of acquired images).

Text-to-image generation holds promise for augmenting limited medical datasets, training AI models, and educational purposes. Studies are actively exploring PEFT for this purpose.¹ For example, the “Prompt to Polyp” research benchmarks two main strategies: fine-tuning large pre-trained latent diffusion models (LDMs) like FLUX and Kandinsky using LoRA, versus training smaller, domain-specific diffusion models like their proposed MSDM from scratch.⁹ These investigations, conducted on datasets such as MedVQA-GI (colonoscopy) and ROCOv2 (radiology images including X-ray, MRI, CT), use LoRA to adapt the large LDMs, assessing factors like LoRA parameters, data augmentation, and data volume on synthesis quality.

For image reconstruction, the PETITE (Parameter Efficient Fine-Tuning for MultI-scanner PET to PET REconstruction) framework leverages a “Mix-PEFT” strategy, combining methods like LoRA, Adapters, SSF (Scaling and Shifting Features), and VPT (Visual Prompt Tuning).³³ PETITE aims to reduce PET scan acquisition times by reconstructing high-quality images from shorter scans, while also addressing variability across different PET scanners. By applying different PEFT techniques to the encoder and decoder components of models like 3D CVT-GAN and UNETR, PETITE achieves performance comparable to full fine-tuning with less than 1% of the trainable parameters, demonstrating significant efficiency gains.³³

3.4. Emerging Applications: Detection, Visual Question Answering (VQA), and Automated Report Generation

PEFT is also facilitating advancements in other critical medical AI applications. While direct applications of PEFT for medical object or lesion detection are less explicitly detailed in the reviewed materials compared to segmentation, many advanced segmentation techniques inherently support or can be adapted for detection.³⁴ For instance, the PRS-Med framework, while focused on segmentation, includes an aspect of tumor identification which is closely related to detection.³⁴

A significant area of growth is in multimodal medical AI, particularly Medical Visual Question Answering (Med-VQA) and automated Medical Report Generation (MRG). The PeFoMed model, for example, employs LoRA to efficiently fine-tune a multimodal architecture (combining the EVA vision encoder with the LLaMA2-chat(7B) LLM) for these tasks.³¹ Using diverse datasets like VQA-RAD, SLAKE, PathVQA (for VQA), and IU-Xray (for MRG), PeFoMed demonstrates high parameter efficiency and achieves performance competitive with, or even outperforming, much larger models like GPT-4v on these specialized medical tasks.³¹

The diverse applications of PEFT across various medical imaging tasks and modalities reveal an important pattern: the optimal PEFT strategy is often not universal. Its effectiveness can be highly dependent on the specific characteristics of the imaging modality (e.g., 2D X-rays versus 3D volumetric CT or MRI scans, noise profiles, resolution) and the nature of the analytical task (e.g., global classification versus fine-grained segmentation or complex generative modeling). For instance, the PETITE framework found that different combinations of PEFT methods (Mix-PEFT) were optimal for the UNETR architecture versus the 3D CVT-GAN model in PET reconstruction, reflecting differences in their underlying components (ViT-based encoders, CNN-based decoders).³³ Similarly, studies on adapting the Segment Anything Model (SAM) indicate that the precise placement of PEFT layers within the Vision Transformer architecture is a critical determinant of success.²⁰ This suggests an emerging necessity for more systematic research to map specific PEFT techniques or their combinations to particular medical imaging challenges. Such research could lead to the development of “PEFT playbooks” or evidence-based guidelines tailored to sub-fields like radiology, pathology, or ophthalmology, assisting researchers and developers in making more informed and effective choices.

Another significant trend illuminated by these applications is PEFT’s role as a crucial bridge for practical multimodal medical AI. Clinical decision-making inherently involves integrating information from diverse sources, most commonly medical images and accompanying textual data such as patient histories or radiology reports. Models like PeFoMed and PRS-Med are leveraging PEFT, particularly LoRA, to adapt very large, general-purpose vision-language foundation models (which are often even more parameter-heavy than unimodal models) for sophisticated medical tasks like VQA, report generation, and reasoning-driven segmentation.³¹ By allowing focused fine-tuning of a small fraction of parameters, often within the language processing or multimodal fusion components, PEFT makes the adaptation of these powerful multimodal architectures computationally feasible. This is accelerating the development of AI systems that can more holistically interpret medical data, potentially leading to tools that better emulate the integrative reasoning processes of human clinicians.

Furthermore, PEFT is demonstrating a dual utility: it not only directly enhances the performance of models on specific clinical tasks (e.g., classification, segmentation) but also indirectly bolsters model capabilities by facilitating efficient data augmentation. As seen in the “Prompt to Polyp” study, PEFT techniques like LoRA are employed to fine-tune generative models (e.g., large diffusion models) for the synthesis of realistic medical images.⁹ These synthetically generated images can then be used to expand limited training datasets. This is particularly valuable in data-scarce medical scenarios, as it can help improve the robustness and generalization of downstream models—which themselves might also be fine-tuned using PEFT. This creates a synergistic cycle: PEFT-enabled image synthesis helps overcome data bottlenecks, which in turn makes PEFT-enabled fine-tuning for specific clinical tasks more effective and reliable. This two-pronged impact of PEFT could be a powerful strategy for bootstrapping high-performing medical AI models, especially in areas where data acquisition is challenging.

Table 2 provides a consolidated summary of key studies applying PEFT in medical imaging, highlighting the methods, tasks, modalities, and principal findings.

Table 2: Summary of Key PEFT Applications in Medical Imaging Studies

4. The Value Proposition: Advantages of PEFT in the Medical Domain

The adoption of PEFT techniques in medical imaging is driven by a compelling value proposition, offering tangible benefits that address many of the inherent challenges in applying large-scale AI to healthcare. These advantages span computational efficiency, data utilization, model management, and knowledge retention.

4.1. Enhanced Parameter and Computational Efficiency

The most immediate and widely recognized advantage of PEFT is the dramatic reduction in the number of trainable parameters compared to full fine-tuning.¹ Typically, PEFT methods update less than 1% to 5% of the total parameters of a large foundation model.¹¹ For instance, LoRA has been shown to save over 99.97% of parameters when adapting models like GPT-3.¹⁶ In medical imaging specific applications, DVPT utilizes only about 0.3-0.5% of the parameters of a Vision Transformer (ViT-B/16) ³⁰, and the PETITE framework for PET reconstruction achieves strong results by training less than 1% of the model parameters.³³

This substantial reduction in trainable parameters directly translates to lower training costs, significantly reduced GPU memory requirements (VRAM), and faster fine-tuning cycles.⁵ Techniques like QLoRA further amplify these benefits by quantizing the frozen base model, drastically cutting down VRAM needs and making it feasible to fine-tune very large models on more modest hardware configurations.²² This efficiency democratizes access to powerful AI, allowing institutions with limited computational budgets to participate in cutting-edge research and development.

4.2. Effectiveness in Low-Data and Resource-Constrained Scenarios

Medical imaging is frequently characterized by “small data” problems, where acquiring large, annotated datasets is difficult and expensive.¹ PEFT methods have demonstrated particular strength in these low-data regimes.¹ By freezing most of the pre-trained model’s weights, PEFT leverages the rich, general-purpose representations learned from vast datasets during pre-training, requiring only a small amount of task-specific data to adapt these representations effectively. Studies have reported performance gains of up to 22% in such scenarios when using PEFT compared to other approaches.¹

The DVPT method, for example, has shown remarkable data efficiency, achieving results comparable to full fine-tuning (which uses 100% of labeled data) with only 40% of the labeled data for 2D medical tasks. This effectively reduces the annotation burden for downstream tasks by approximately 60% without a significant drop in performance.³⁰ This capability is invaluable in medicine, enabling the development of effective AI tools even when large task-specific datasets are unattainable.

4.3. Improved Storage, Modularity, and Model Reusability

Full fine-tuning necessitates storing an entire copy of the large foundation model for each specific task or dataset. In contrast, PEFT requires storing only the small set of task-specific PEFT parameters (e.g., LoRA matrices, adapter weights, or prompt vectors) alongside a single shared copy of the frozen pre-trained model.⁵ This leads to massive savings in storage costs. For instance, DVPT can reduce the storage cost for a ViT-B/16 model by up to 99%.³⁰

Furthermore, many PEFT methods, such as Adapter Tuning, are inherently modular.⁵ Task-specific adapters can be treated as plug-and-play modules that can be easily inserted, removed, or swapped. This allows a single, large pre-trained backbone to be efficiently reused across numerous downstream tasks by simply loading the relevant lightweight adapter, greatly simplifying model management and deployment in diverse clinical workflows.

4.4. Potential for Better Transferability and Mitigating Catastrophic Forgetting

By keeping the majority of the pre-trained model’s weights frozen, PEFT methods are better at preserving the general knowledge and robust representations learned during the initial large-scale pre-training phase.⁴ This can lead to improved transferability to new, related tasks or datasets. For example, prefix-tuning has been noted for its ability to preserve the model’s representation space effectively.⁴

This preservation of pre-trained knowledge also helps in mitigating “catastrophic forgetting,” a common problem in neural networks where a model forgets previously learned information when sequentially trained on new tasks.⁵ Since PEFT modifies only a small portion of the model for each new task, the core knowledge remains largely intact, facilitating more stable continual learning. Additionally, some studies suggest that PEFT can retain a model’s robustness to distribution shifts better than full fine-tuning, which is critical for medical applications where data can vary significantly across populations or acquisition settings.¹⁴

Beyond direct deployment, an often-overlooked benefit of PEFT is its impact on the pace of research and development in medical AI. The significantly faster fine-tuning cycles enabled by PEFT allow researchers to conduct more experiments, test more hypotheses, and iterate on model designs more rapidly within the same timeframe or computational budget.⁵ Medical AI research inherently involves extensive experimentation with different model architectures, hyperparameter configurations, and dataset variations. By reducing the turnaround time for these experimental cycles, PEFT indirectly accelerates the overall rate of innovation. This allows for quicker exploration of novel approaches and more efficient refinement of existing ones, potentially leading to more rapid breakthroughs in the application of AI to solve pressing medical challenges.

5. Navigating the Hurdles: Challenges and Limitations of PEFT in Medical Imaging

Despite the significant advantages and growing adoption of PEFT in medical imaging, several challenges and limitations must be acknowledged and addressed to ensure its responsible and effective deployment. These range from technical considerations regarding model performance and generalization to broader ethical and safety concerns.

5.1. Ensuring Robust Generalization and Avoiding Overfitting

While PEFT methods are often lauded for their data efficiency, the challenge of ensuring that models generalize robustly from limited medical data to unseen patient cases, diverse demographic populations, and variations in imaging equipment remains critical.¹ Medical images can exhibit subtle but clinically significant variations, and models adapted with PEFT must be able to capture these nuances without overfitting to the specific characteristics of the training set. The risk of overfitting might persist, albeit potentially in a different form, as the model could over-specialize the small set of tunable parameters to the training data, especially if this data is not sufficiently representative of the target population or contains inherent biases.

5.2. Addressing Domain Shift and Data Heterogeneity

Medical data is notoriously heterogeneous, arising from differences in imaging scanners, acquisition protocols, reconstruction algorithms, patient demographics, and even the subtle variations in how diseases manifest across individuals.⁵³ This phenomenon, known as domain shift, can lead to a significant degradation in model performance when a model trained in one domain (e.g., on data from one hospital or scanner type) is applied to another. PEFT-adapted models must demonstrate robustness to such shifts. Studies like the PETITE framework for multi-scanner PET reconstruction explicitly tackle this issue, highlighting the importance of developing PEFT strategies that can account for inter-scanner variability.³³ Ensuring that the limited parameter updates in PEFT are sufficient to adapt to these diverse data distributions is an ongoing research area.

5.3. Maintaining Performance Fidelity Compared to Full Fine-Tuning

Although many PEFT techniques aim to achieve performance comparable to full fine-tuning (FT), and sometimes succeed, especially in low-data scenarios, there can still be an accuracy gap in certain situations.⁶ This gap might become more apparent for highly complex medical tasks that require nuanced understanding or when the number of tuned parameters is extremely restricted. The choice of PEFT method, its specific configuration (e.g., rank in LoRA, adapter bottleneck dimension), and the layers to which it is applied are all critical factors that can influence the final performance. For instance, the DoRA technique was proposed partly to address the learning capacity differences observed between LoRA and FT.²⁴ Continuous efforts are needed to refine PEFT methods to consistently match or exceed FT performance across a broader range of applications without compromising efficiency.

5.4. Complexity of Choosing and Optimizing PEFT Strategies

The rapid proliferation of diverse PEFT methods, each with its own set of hyperparameters and architectural considerations, presents a new layer of complexity for researchers and practitioners. Selecting the optimal PEFT strategy for a given medical imaging task, foundation model, and set of resource constraints is not straightforward (as discussed in Insight 2.1). Furthermore, the hyperparameter tuning for the PEFT modules themselves (e.g., learning rates for adapters, rank for LoRA, prompt length for prompt tuning) can be as involved as tuning a smaller model from scratch, requiring careful experimentation and validation. This complexity can be a barrier to adoption if not accompanied by clear guidelines or automated selection tools.

5.5. Ethical Considerations, Safety, and Privacy Implications

The application of AI in medicine, regardless of the fine-tuning method, carries significant ethical responsibilities. PEFT, while offering efficiency, does not inherently resolve these concerns and may introduce new nuances.

• Reliability and Trustworthiness: It is paramount that medical AI models, including those adapted with PEFT, are highly reliable, well-calibrated (i.e., their confidence scores reflect true likelihoods), and do not perpetuate or amplify biases present in the training data.⁷ The safety of patients depends on the accuracy and dependability of these tools.
• Factual Accuracy and Hallucinations: For generative PEFT applications, such as LLMs adapted for medical report generation or VQA, ensuring factual accuracy and preventing the generation of misleading or harmful “hallucinations” is a critical safety concern.⁵⁴
• Patient Privacy: While PEFT reduces the need to transfer or store massive datasets for fine-tuning each model instance, the adaptation process itself still involves training on sensitive patient data.¹ Robust privacy-preserving techniques, such as federated learning (where PEFT’s small update size is advantageous), differential privacy, or secure multi-party computation, must be employed to protect patient confidentiality.
• Opacity and Accountability: In complex deployment scenarios like federated learning, which might utilize PEFT, issues such as “federation opacity” (lack of transparency into the data and processes at each participating site) can create a “double black box problem,” making it difficult to understand model behavior, debug errors, or assign accountability.⁵⁵

The drive for efficiency through PEFT, by its nature of tuning fewer parameters, introduces a nuanced tension when juxtaposed with the stringent demands for trustworthiness in medical AI.³ While limiting parameter updates makes models more accessible and quicker to adapt, it might, in certain complex medical scenarios, constrain the model’s ability to fully internalize all the subtle, safety-critical nuances of a task or to comprehensively unlearn biases inherited from its general pre-training. This is not to say PEFT is inherently less safe, but rather that the very mechanism of restricted adaptation warrants careful scrutiny. The evaluation of PEFT methods in medicine must therefore extend far beyond standard accuracy metrics. It necessitates rigorous assessments of model robustness against data perturbations, calibration of predictive uncertainties, detection and mitigation of algorithmic bias, and thorough analysis of failure modes, particularly those with potential clinical impact. The significant efficiency gains afforded by PEFT should not inadvertently lead to a compromise in the depth of adaptation required for ensuring the utmost safety and fairness in sensitive medical applications. This underscores a pressing need for the development of specialized evaluation protocols and benchmarks specifically designed for PEFT in safety-critical medical domains, potentially incorporating adversarial testing, counterfactual analysis, and enhanced explainability metrics to build confidence in these efficiently adapted models.

Furthermore, PEFT interacts with the long-standing “data scarcity versus model capacity” dilemma in medical AI. Foundation models are, by definition, high-capacity architectures, capable of learning complex patterns from vast datasets.¹ Medical datasets, conversely, are often limited in size. PEFT attempts to bridge this gap by fine-tuning only a small portion of this large capacity using the available medical data. While this strategy effectively mitigates overfitting by constraining the number of tunable parameters, it also implies that a substantial part of the model’s learned capacity—derived from general-domain data during pre-training—remains largely “untouched” or less directly shaped by the specific medical task at hand. This raises an important question: how much of this preserved pre-trained knowledge is genuinely beneficial and transferable to highly specialized medical tasks, versus how much might be irrelevant or even potentially interfering? While PEFT offers an efficient adaptation pathway, it’s conceivable that for certain niche medical features or extremely subtle diagnostic cues, more extensive parameter changes than typically allowed by PEFT might be necessary for the model to fully learn and represent them. This points to a need for ongoing research into the intricate interplay between the volume and nature of pre-trained knowledge, the scope and type of parameters modified by PEFT techniques, and the specificity and complexity of the target medical task. Such investigations could pave the way for more adaptive PEFT methods that can dynamically adjust the extent or focus of fine-tuning based on an assessment of task complexity and data availability, ensuring an optimal balance between efficiency and adaptive capacity.

6. Spotlight: Noteworthy PEFT Frameworks and Models in Medical AI

The rapid evolution of PEFT has led to the development of several specialized frameworks and models tailored for, or demonstrating significant impact in, the medical imaging domain. These innovations often combine foundational PEFT principles with domain-specific insights to address unique challenges in healthcare AI.

• PETITE (Parameter Efficient Fine-Tuning for MultI-scanner PET to PET REconstruction) and Mix-PEFT: This framework is designed to tackle challenges in Positron Emission Tomography (PET) image reconstruction, specifically addressing variability across different scanners and aiming to reduce scan acquisition times.³³ PETITE employs a Mix-PEFT strategy, which involves applying different PEFT methods (including LoRA, Adapters, SSF, and VPT) independently to the encoder and decoder components of underlying reconstruction models like 3D CVT-GAN and UNETR. Key findings indicate that PETITE can achieve performance comparable to full fine-tuning while using less than 1% of the trainable parameters. For instance, the optimal Mix-PEFT configuration for the 3D CVT-GAN model was found to be VPT in the encoder and LoRA in the decoder, whereas for the UNETR model, LoRA in the encoder and SSF in the CNN-based decoder proved most effective. This demonstrates significant parameter and computational efficiency, crucial for practical PET imaging.
• DVPT (Dynamic Visual Prompt Tuning): DVPT is a novel PEFT method that has shown strong results in both medical image classification (e.g., diabetic retinopathy grading from retinal images) and segmentation across various modalities (e.g., polyps in colonoscopy, skin lesions, cardiac MRI, multi-organ CT).³⁰ It introduces trainable prompt tokens at the input layer and a Cross-Attention Visual Prompt Tuning (CAVPT) module within the Transformer blocks. This module allows prompts to dynamically query and aggregate sample-specific features. DVPT is characterized by its high parameter efficiency (tuning only ~0.3-0.5% of a ViT-B/16 model’s parameters) and remarkable data efficiency, reportedly reducing labeled data requirements by up to 60% while outperforming other PEFT methods and sometimes even full fine-tuning.
• EPT (Embedded Prompt Tuning): EPT is proposed to enhance few-shot medical image classification by improving how prompt tokens are integrated into Transformer architectures.³⁵ It achieves this by embedding prompt tokens into the expanded channels of the model. EPT has been shown to outperform several state-of-the-art fine-tuning methods in scenarios with limited medical training data. A conceptual contribution of this work is the perspective that prompt tuning can act as a “distribution calibrator,” helping to mitigate anomalies in the feature space of foundation models.
• PeFoMed (Parameter Efficient Fine-tuning of Multimodal Large Language Models for Medical Imaging): PeFoMed focuses on adapting Multimodal Large Language Models (MLLMs), specifically an architecture combining the EVA vision encoder with the LLaMA2-chat(7B) LLM, for complex medical tasks like Visual Question Answering (Med-VQA) and automated Medical Report Generation (MRG).³¹ It utilizes LoRA applied to the LLM component and a vision projection layer. With only 56.63 million trainable parameters (compared to the 7 billion parameters of a model like LLaVA-Med), PeFoMed achieves high parameter efficiency. It has demonstrated performance competitive with specialized models and has even outperformed general large multimodal models like GPT-4v on specific medical benchmarks. The study also highlights the benefit of a two-stage fine-tuning strategy (initial fine-tuning on image captioning, followed by task-specific fine-tuning for VQA/MRG).
• PEMMA (Parameter-Efficient Multi-Modal Adaptation for Medical Image Segmentation): The PEMMA framework targets the adaptation of segmentation models across different imaging modalities and for incorporating additional data types.³² It uses LoRA or DoRA to efficiently update attention weights, enabling a model pre-trained on CT scans, for example, to be adapted for PET scan segmentation or for prognostic tasks using Electronic Health Record (EHR) data. Applied to head and neck cancer segmentation using the HECKTOR dataset, PEMMA achieved Dice scores comparable to early fusion techniques but with only 8% of the trainable parameters, and notably improved performance on PET scans. It also shows promise for continual learning scenarios.
• TPP (Target Parameter Pre-training): TPP is a framework designed to enhance the effectiveness of various PEFT methods by improving the initialization of the newly introduced trainable parameters (referred to as “target parameters”).⁸ Instead of random initialization, TPP introduces an additional pre-training stage specifically for these target parameters (e.g., LoRA matrices, adapter weights) using the downstream task’s training data, while keeping the main backbone frozen. This allows the target parameters to learn dataset-specific representations before the main PEFT fine-tuning process. TPP is presented as a plug-and-play solution that consistently improves the fine-tuning performance of various PEFT methods across multiple medical datasets and modalities.
• PRS-Med (Position Reasoning Segmentation with Vision-Language Model in Medical Imaging): PRS-Med is a framework for medical image segmentation that uniquely incorporates position reasoning capabilities, enabling the model to not only segment regions like tumors but also explain their location.³⁴ It achieves this by using LoRA to fine-tune a Vision-Language Model (LLaVA-Med) in conjunction with a TinySAM (Segment Anything Model) image encoder. PRS-Med has demonstrated strong performance in both segmentation accuracy and position reasoning across a diverse range of six imaging modalities, including CT, MRI, X-ray, ultrasound, endoscopy, and standard RGB images.
• QLoRA Applications in Medical LLMs: QLoRA is being increasingly applied to fine-tune Large Language Models for various medical decision support applications.²⁰ For instance, systems are being developed that integrate LLMs (like Llama 3.2-3B-Instruct) with Retrieval-Augmented Generation (RAG) from hospital-specific databases (EHRs, clinical guidelines) and then fine-tune these LLMs using QLoRA.²¹ This approach allows for enhanced disease prediction, treatment suggestions, and efficient summarization of complex medical reports. QLoRA’s extreme memory optimization enables such fine-tuning on relatively modest hardware, making it practical for smaller healthcare institutions to adapt LLMs to their local clinical practices and data, while aiming to preserve the integrity of medical information. QLoRA is also being used to adapt Vision Transformers for segmentation tasks.²⁰

The development of these specialized frameworks reveals a clear trend towards more sophisticated and hybrid PEFT strategies in medical imaging. Rather than relying on a single, off-the-shelf PEFT technique, advanced applications are increasingly adopting modular and composable approaches. Frameworks like PETITE’s Mix-PEFT explicitly combine different PEFT methods, applying them strategically to distinct components of a model architecture (e.g., VPT for an encoder, LoRA for a decoder) to maximize synergy.³³ The TPP framework acts as a preparatory stage that enhances the performance of subsequent PEFT methods by providing better-initialized target parameters.⁸ Similarly, PRS-Med integrates LoRA with specific, powerful model components like LLaVA-Med and SAM to achieve its reasoning-segmentation capabilities.³⁴ The QLoRA technique itself is an inherent combination of LoRA and quantization.²¹ This indicates that the future of PEFT in complex domains like medicine may lie in developing intelligent strategies for combining the strengths of various efficient tuning and model compression methods, tailored to the specific demands of the model architecture and the medical task. This could spur research into the theoretical underpinnings of how different PEFT methods interact and how to best compose them for optimal results, potentially leading to automated methods for discovering such compositions, akin to the concept of AUTOPEFT.³

Furthermore, these innovative frameworks highlight PEFT’s crucial role in facilitating the creation and adaptation of domain-specific multimodal foundation models for medicine. Models like PeFoMed and PRS-Med leverage PEFT (primarily LoRA) to efficiently fine-tune large, general-purpose vision-language architectures (such as LLaMA or LLaVA-Med derivatives) to the unique requirements of the medical domain.³¹ These adaptations enable tasks that inherently require the understanding and generation of both visual information (from medical images) and textual information (clinical questions, diagnostic reports, reasoning explanations). PEFT makes this process tractable by focusing the adaptation on a small subset of parameters, often within the language processing units or the multimodal fusion layers. This trend suggests that PEFT is not merely a tool for unimodal task adaptation but a critical component in the development pipeline for sophisticated, domain-specific AI systems capable of handling the inherently multimodal nature of medical data and clinical workflows. We may witness the emergence of a new generation of “PEFT-ed” medical foundation models that are both highly capable and efficient enough for widespread clinical deployment, potentially even fine-tuned with institution-specific data to align with local practices, as explored in QLoRA-based LLM applications.²¹

7. The Path Forward: Future Research Directions and Concluding Remarks

Parameter-Efficient Fine-Tuning has undeniably opened new frontiers for applying advanced AI in medical imaging. However, the journey is far from over. Several exciting and critical research directions lie ahead, promising to further enhance the capabilities, applicability, and trustworthiness of PEFT in the medical domain.

7.1. Opportunities for Novel PEFT Methods Tailored for Medical Data

While existing PEFT methods have shown considerable success, there is ample room for developing novel techniques specifically designed for the unique characteristics of medical data.⁴⁸ Future research could explore:

• Domain-Knowledge Integration: PEFT methods that can explicitly incorporate medical domain knowledge, such as anatomical atlases, physiological models, or known imaging physics. This could guide the fine-tuning process to learn more clinically relevant features.
• Handling Complex Medical Data Structures: Developing PEFT techniques optimized for 3D and 4D medical data (e.g., volumetric CT/MRI, dynamic imaging sequences), graph-based medical data (e.g., patient networks, molecular structures), and other complex, structured or unstructured medical information.
• Robustness to Artifacts and Noise: Designing PEFT strategies that are inherently more robust to common image artifacts, noise, and variations in image quality prevalent in real-world medical imaging.⁷ This could involve PEFT modules that learn to identify and mitigate the impact of such imperfections.

7.2. The Role of PEFT in Federated Learning and Privacy-Preserving AI

The small parameter footprint of PEFT makes it exceptionally well-suited for Federated Learning (FL) scenarios.⁵⁵ In FL, multiple institutions collaboratively train a model without sharing their raw patient data; instead, only model updates (or PEFT parameters in this case) are exchanged. This significantly enhances patient privacy, a paramount concern in healthcare. Future work should focus on:

• Optimizing PEFT techniques for efficient aggregation and communication in FL settings.
• Addressing challenges specific to FL with PEFT, such as statistical heterogeneity across participating sites and ensuring fairness.
• Investigating how to effectively navigate issues like “federation opacity” ⁵⁵ when PEFT is employed within FL frameworks to maintain transparency and accountability.

7.3. Integrating PEFT with Continual Learning Paradigms

Medical knowledge and clinical practices are constantly evolving, and new patient data becomes available over time. Medical AI models must be able to adapt to these changes through continual learning (CL) without catastrophically forgetting previously learned information.⁵ PEFT’s ability to preserve pre-trained knowledge by freezing most parameters makes it a natural fit for CL. The PEMMA framework, for instance, has already shown promise in supporting continual learning.³² Research should explore:

• Developing PEFT-based CL strategies that allow models to seamlessly integrate new medical knowledge and adapt to data drift over time.
• Mechanisms to dynamically allocate or expand PEFT parameters as new tasks or data distributions are encountered.

7.4. Optimizing PEFT for Resource-Constrained Hardware and Edge Devices

The push to bring AI capabilities to the point of care necessitates deploying models on resource-constrained hardware, such as portable medical devices, embedded systems in imaging equipment, or local clinic servers.⁵ Future PEFT research should continue to prioritize:

• Extreme model compression techniques (e.g., more aggressive quantization, sophisticated pruning methods) tightly integrated with PEFT.
• Developing PEFT methods that are not only parameter-efficient but also computationally efficient during both the fine-tuning and inference stages on low-power processors.

7.5. Standardization of Benchmarks and Evaluation Protocols

To ensure rigorous and fair comparison of different PEFT methods in the medical imaging context, there is a pressing need for standardized benchmarks and evaluation protocols.¹ The initiative by Dutt et al. to create a structured benchmark for PEFT in medical image analysis is a step in this direction.¹ Future efforts should aim to:

• Establish comprehensive benchmark datasets covering diverse medical tasks, imaging modalities, and patient populations.
• Develop and adopt evaluation metrics that capture not only predictive accuracy but also computational efficiency, model robustness, fairness, calibration, and, crucially, clinical utility and interpretability.

7.6. Addressing Safety, Interpretability, and Trustworthiness

Ultimately, the success of PEFT-adapted AI models in medicine hinges on their safety, interpretability, and the trust they inspire in clinicians and patients.⁷ Ongoing research must focus on:

• Ensuring that PEFT models are reliable, robust to adversarial attacks or unexpected inputs, and aligned with human-defined clinical objectives and ethical guidelines.
• Developing techniques to enhance the interpretability of PEFT-adapted models, allowing clinicians to understand the basis of their predictions or recommendations.
• Mitigating biases and ensuring fairness across different demographic groups.

The convergence of PEFT, Federated Learning (FL), and Continual Learning (CL) appears to be a particularly potent combination for the future of dynamic medical AI. The inherent parameter efficiency of PEFT provides the compact model updates essential for privacy-preserving FL and for the iterative adaptation required by CL.⁵⁵ Medical environments are characterized by their dynamic nature—new data is constantly generated, clinical guidelines evolve, and patient populations shift—and by the critical need to protect sensitive data, which often resides in distributed silos. PEFT offers an efficient mechanism for model updates; FL provides a framework for collaborative training on these distributed datasets without centralizing raw data; and CL equips models with the ability to learn from new information over time while retaining past knowledge. This synergy suggests that the next generation of robust, adaptable, and privacy-respecting medical AI systems will likely arise from architectures that tightly integrate these three paradigms. For example, a consortium of hospitals could collaboratively and continually fine-tune a shared foundation model using PEFT within a federated learning setup, allowing the model to benefit from diverse data sources while adapting to the latest medical insights. This points to a significant research frontier focused on developing algorithms and frameworks that effectively unify PEFT, FL, and CL, while also addressing their combined challenges, such as managing statistical heterogeneity in FL, preventing catastrophic forgetting in CL, and designing efficient parameter aggregation strategies for PEFT-based updates in distributed settings.

Reference:

1. arxiv.org, accessed on June 5, 2025, https://arxiv.org/html/2305.08252v4
2. [2305.08252] Parameter-Efficient Fine-Tuning for Medical Image Analysis: The Missed Opportunity – arXiv, accessed on June 5, 2025, https://arxiv.org/abs/2305.08252
3. arxiv.org, accessed on June 5, 2025, https://arxiv.org/pdf/2410.19878
4. arxiv.org, accessed on June 5, 2025, https://arxiv.org/html/2411.00029v1
5. Advances in Parameter-Efficient Fine-Tuning: Optimizing Foundation Models for Scalable AI, accessed on June 5, 2025, https://www.preprints.org/manuscript/202503.2048/v1
6. Revisiting Fine-Tuning: A Survey of Parameter-Efficient Techniques for Large AI Models – Preprints.org, accessed on June 5, 2025, https://www.preprints.org/manuscript/202504.0743/v1/download
7. Physical foundations for trustworthy medical imaging: a review for artificial intelligence researchers – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2505.02843v1
8. Parameter-Efficient Fine-Tuning for Medical Image Analysis via Target Parameter Pre-training – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2408.15011v1
9. Prompt to Polyp: Clinically-Aware Medical Image Synthesis with Diffusion Models – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2505.05573v1
10. (PDF) Prompt to Polyp: Clinically-Aware Medical Image Synthesis with Diffusion Models, accessed on June 5, 2025, https://www.researchgate.net/publication/391658122_Prompt_to_Polyp_Clinically-Aware_Medical_Image_Synthesis_with_Diffusion_Models
11. arxiv.org, accessed on June 5, 2025, https://arxiv.org/pdf/1902.00751
12. Parameter-Efficient Fine-Tuning for Large Models: A Comprehensive Survey – arXiv, accessed on June 5, 2025, https://arxiv.org/pdf/2403.14608?
13. MCML – Research Group Zeynep Akata – Munich Center for Machine Learning, accessed on June 5, 2025, https://mcml.ai/research/groups/akata/
14. Lessons and Insights from a Unifying Study of Parameter-Efficient Fine-Tuning (PEFT) in Visual Recognition – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2409.16434v5
15. Parameter-Efficient Fine-Tuning for Pre-Trained Vision Models: A Survey – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2402.02242v4
16. Parameter-Efficient Fine-Tuning for Foundation Models – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2501.13787v1
17. arxiv.org, accessed on June 5, 2025, https://arxiv.org/pdf/2501.03291
18. arxiv.org, accessed on June 5, 2025, https://arxiv.org/pdf/2101.00190
19. arxiv.org, accessed on June 5, 2025, https://arxiv.org/abs/2504.07448
20. Parameter Efficient Fine-Tuning of Segment Anything Model for Biomedical Imaging – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2502.00418v2
21. Lightweight Clinical Decision Support System using QLoRA-Fine-Tuned LLMs and Retrieval-Augmented Generation – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2505.03406v1
22. Lightweight Clinical Decision Support System using QLoRA-Fine-Tuned LLMs and Retrieval-Augmented Generation – arXiv, accessed on June 5, 2025, https://www.arxiv.org/pdf/2505.03406
23. [2402.18039] ResLoRA: Identity Residual Mapping in Low-Rank Adaption – arXiv, accessed on June 5, 2025, https://arxiv.org/abs/2402.18039
24. DoRA: Weight-Decomposed Low-Rank Adaptation – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2402.09353v3
25. [2402.09353] DoRA: Weight-Decomposed Low-Rank Adaptation – arXiv, accessed on June 5, 2025, https://arxiv.org/abs/2402.09353
26. [2401.16137] X-PEFT: eXtremely Parameter-Efficient Fine-Tuning for Extreme Multi-Profile Scenarios – arXiv, accessed on June 5, 2025, https://arxiv.org/abs/2401.16137
27. [2502.01033] PARA: Parameter-Efficient Fine-tuning with Prompt Aware Representation Adjustment – arXiv, accessed on June 5, 2025, https://arxiv.org/abs/2502.01033
28. [2101.00190] Prefix-Tuning: Optimizing Continuous Prompts for Generation – ar5iv – arXiv, accessed on June 5, 2025, https://ar5iv.labs.arxiv.org/html/2101.00190
29. Revisiting Prefix-tuning: Statistical Benefits of Reparameterization among Prompts – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2410.02200v5
30. arxiv.org, accessed on June 5, 2025, https://arxiv.org/abs/2307.09787
31. arxiv.org, accessed on June 5, 2025, https://arxiv.org/html/2401.02797v2
32. Efficient Parameter Adaptation for Multi-Modal Medical Image Segmentation and Prognosis – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2504.13645v1
33. papers.miccai.org, accessed on June 5, 2025, https://papers.miccai.org/miccai-2024/paper/0560_paper.pdf
34. PRS-Med: Position Reasoning Segmentation with Vision-Language Model in Medical Imaging – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2505.11872v2
35. [2407.01003] Embedded Visual Prompt Tuning – arXiv, accessed on June 5, 2025, https://arxiv.org/abs/2407.01003
36. Parameter-Efficient Fine-Tuning for Medical Image Analysis: The Missed Opportunity, accessed on June 5, 2025, https://openreview.net/forum?id=LVRhXa0q5r
37. arxiv.org, accessed on June 5, 2025, https://arxiv.org/pdf/2305.08252
38. arxiv.org, accessed on June 5, 2025, https://arxiv.org/pdf/2407.01003
39. arxiv.org, accessed on June 5, 2025, https://arxiv.org/pdf/2502.00418
40. Computer Science Apr 2025 – arXiv, accessed on June 5, 2025, http://arxiv.org/list/cs/2025-04?skip=6325&show=2000
41. Machine Learning Apr 2025 – arXiv, accessed on June 5, 2025, http://arxiv.org/list/cs.LG/2025-04?skip=2500&show=1000
42. arxiv.org, accessed on June 5, 2025, https://arxiv.org/pdf/2504.13645
43. PRS-Med: Position Reasoning Segmentation with Vision-Language Model in Medical Imaging – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2505.11872v1
44. Xuchen-Li/llm-arxiv-daily: Automatically update arXiv papers about LLM Reasoning, LLM Evaluation, LLM & MLLM and Video Understanding using Github Actions. – GitHub, accessed on June 5, 2025, https://github.com/Xuchen-Li/llm-arxiv-daily
45. arxiv.org, accessed on June 5, 2025, https://arxiv.org/pdf/2505.11872
46. arxiv.org, accessed on June 5, 2025, https://arxiv.org/pdf/2505.05573
47. Transformers in medical image analysis | Intelligent Medicine – MedNexus, accessed on June 5, 2025, https://mednexus.org/doi/abs/10.1016/j.imed.2022.07.002
48. Prospects of deep learning for medical imaging – pfm :: Precision and Future Medicine, accessed on June 5, 2025, https://www.pfmjournal.org/m/journal/view.php?number=32
49. Rethinking Boundary Detection in Deep Learning-Based Medical Image Segmentation, accessed on June 5, 2025, https://arxiv.org/html/2505.04652v1
50. arxiv.org, accessed on June 5, 2025, https://arxiv.org/pdf/2401.02797
51. Pre-training Everywhere: Parameter-Efficient Fine-Tuning for Medical Image Analysis via Target Parameter Pre-training – ResearchGate, accessed on June 5, 2025, https://www.researchgate.net/publication/383461067_Pre-training_Everywhere_Parameter-Efficient_Fine-Tuning_for_Medical_Image_Analysis_via_Target_Parameter_Pre-training
52. arxiv.org, accessed on June 5, 2025, https://arxiv.org/pdf/2408.15011
53. [2403.12167] A Systematic Review of Generalization Research in Medical Image Classification – arXiv, accessed on June 5, 2025, https://arxiv.org/abs/2403.12167
54. Safety challenges of AI in medicine in the era of large language models – arXiv, accessed on June 5, 2025, https://arxiv.org/html/2409.18968v2
55. Federated learning, ethics, and the double black box problem in medical AI – PhilArchive, accessed on June 5, 2025, https://philarchive.org/archive/HATFLE
56. Foundation Models in Radiology: What, How, Why, and Why Not – RSNA Journals, accessed on June 5, 2025, https://pubs.rsna.org/doi/10.1148/radiol.240597