Curriculum Learning in 3D Medical Imaging: Advancing Diagnostic and Therapeutic Applications

Curriculum learning, a machine learning paradigm inspired by human cognitive development, involves training models on examples of progressively increasing difficulty. 3D medical imaging, encompassing modalities such as Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and Positron Emission Tomography (PET), offers unparalleled anatomical and functional insights, revolutionizing diagnosis and treatment planning. However, its full potential in artificial intelligence (AI) is often constrained by challenges inherent to medical data, including scarcity, high annotation costs, class imbalance, and label noise.

The report details how CL effectively addresses these critical limitations, leading to significant improvements in model performance, faster convergence, and enhanced generalizability across diverse 3D medical imaging tasks, notably segmentation, classification, and registration. Key advancements discussed include the evolution from static to dynamic, uncertainty-driven difficulty metrics, the emergence of computationally efficient implicit curricula, and CL’s crucial role in fine-tuning large foundation models for specialized medical applications.

Curriculum Learning: Principles and Core Concepts

Curriculum learning is a sophisticated machine learning technique that structures the training process by presenting examples to a model in an order of increasing difficulty. This methodology is deeply inspired by human cognitive development, where individuals learn by first mastering simple concepts before progressing to more complex ones.

The successful application of CL relies on two critical elements: a difficulty measurer, which quantifies the inherent “hardness” of each training example, and a learning scheduler, which dictates the sequence and pace at which these examples are introduced to the model during training^. There are several major variations in how this technique is applied. The concept of “difficulty” can be externally imposed, for instance, through human expert annotations or predefined heuristics; for example, in language modeling, shorter sentences might be classified as easier than longer ones. Another approach is to use the performance of another model, with examples accurately predicted by that model being classified as easier, providing a connection to boosting.

The progression of difficulty can be implemented steadily and continuously, or in distinct, discrete epochs. The schedule might be deterministic, following a fixed progression, or probabilistic, introducing examples based on a probability distribution. Simple fixed schedules might involve training on easy examples for a set number of iterations before exposing the model to the full dataset. More advanced approaches, known as self-paced learning, dynamically adjust the increase in difficulty in proportion to the model’s current performance and mastery of the training set. As CL primarily focuses on the strategic selection and ordering of training data, it is highly versatile and can be seamlessly integrated with a multitude of other machine learning techniques. Furthermore, some interpretations of CL extend to progressively increasing the complexity of the model itself, such as gradually expanding the number of model parameters.

However, medical imaging data is often limited and noisy 7, and is difficult and expensive to acquire.8 This inherent data scarcity and imperfection in 3D medical imaging create a critical needs for techniques like CL. Curriculum learning, by intelligently ordering, weighting, or sampling data ⁸, allows models to learn effectively from less data or data with imperfections such as noise or class imbalance. Consequently, CL is not merely an enhancement; it is a necessary enabler for deep learning to effectively exploit the diagnostic richness of 3D medical images in real-world clinical scenarios. Without CL, the full diagnostic and analytical potential of 3D imaging for AI applications would remain largely untapped due to the “annotation bottleneck”.⁹ This dependency positions CL as a cornerstone for the practical scalability and clinical translation of AI in 3D medical imaging. It suggests that continued investment in CL research is paramount for overcoming the fundamental data challenges that hinder widespread AI adoption in healthcare, ensuring that the computational power of deep learning can fully unlock the diagnostic value embedded in complex 3D medical datasets.

Foundational Principles of Curriculum Learning

Mechanisms of Difficulty Definition and Data Ordering

A prerequisite for any curriculum learning approach is the establishment of a quantifiable and meaningful measure of “difficulty” for each training example. This definition can originate from various sources. Difficulty can be pre-defined through human annotation or by applying external heuristics. For instance, in natural language processing, shorter sentences are often heuristically classified as easier. In medical imaging, prior domain knowledge can be leveraged, such as keywords extracted from medical reports, the frequency of specific sample types, established medical classification standards, or even quantifying inconsistencies in expert annotations.⁸ This approach is often termed “static CL” when curriculum probabilities are predetermined.⁸

An alternative method involves assessing the performance of a model itself. Examples that are accurately and confidently predicted by the model are then classified as “easier”. This provides a dynamic, model-driven feedback loop for difficulty assessment. A sophisticated and increasingly popular approach, particularly when explicit domain knowledge is scarce, is to dynamically quantify difficulty by measuring the model’s uncertainty in its predictions. Techniques like Monte-Carlo (MC) dropout can be employed to estimate the predictive entropy of the model’s output. Samples associated with low predictive entropy (i.e., high certainty) are considered “easier,” while those with high entropy (high uncertainty) are “harder”.⁸ This dynamic approach prioritizes samples with high information gain at the early stages of training, allowing the model to rapidly reduce errors on highly misleading or uncertain examples.⁸

Once difficulty is defined, CL employs various strategies to order and present the training data effectively. Instead of the typical random permutation of training samples in each epoch, CL intelligently reorders the dataset. This “smart” probabilistic ordering ensures that easier examples are presented earlier, gradually introducing more challenging cases at later stages of training.⁸ Another strategy involves starting with a smaller, easier subset of the training data and progressively increasing the subset size over epochs. The primary purpose is to mitigate the adverse effects of outliers or highly difficult examples during the initial, sensitive phases of training.⁸ Additionally, individual scalar weights can be assigned to training samples based on their calculated curriculum probabilities. These weights are then applied to the classification loss, effectively reducing the contribution to the loss from samples deemed to have low priority, such as very easy or potentially noisy samples that might hinder early learning.⁸

The evolution of difficulty definition from external, often subjective, heuristics to internal, dynamic, and uncertainty-driven metrics is a significant development. Early curriculum learning relied on predefined notions of difficulty, which were inherently limited by the cost and subjectivity of human expertise ⁸ and might not perfectly reflect what the model itself finds challenging. The progression to dynamic, model-driven difficulty metrics, particularly those based on the model’s own uncertainty in its predictions, using techniques like Monte-Carlo dropout to estimate predictive entropy ⁸ or Dirichlet distribution classifiers ¹⁰, is crucial. This shift makes curriculum learning more scalable and adaptive, as it reduces the reliance on expensive and time-consuming expert input for curriculum design.⁸ By allowing the model to self-assess its learning progress and identify areas of high uncertainty, which often correspond to difficult or informative examples, the curriculum becomes intrinsically aligned with the model’s current learning state, leading to more efficient and effective training, especially in the presence of noisy or imbalanced medical data. This trend suggests that future CL research will increasingly focus on developing more sophisticated, autonomous “curriculum agents” capable of learning and adapting optimal difficulty metrics and pacing functions without human intervention. This could involve integrating reinforcement learning to dynamically discover the most effective learning paths, further enhancing the robustness and applicability of CL in complex, real-world medical AI systems.

Scheduling Strategies: Fixed vs. Self-Paced Learning

Scheduling strategies in curriculum learning can be broadly categorized. Fixed schedules are straightforward approaches where the curriculum progression is predetermined and static. For example, a model might be trained exclusively on easy examples for the first half of its available training iterations, and then on the entire dataset, including harder examples, for the second half.¹ This aligns with the concept of “static CL,” where the curriculum probabilities are set a priori.⁸

Representing a more dynamic and adaptive approach, self-paced learning adjusts the increase in difficulty proportionally to the model’s current performance on the training set.¹ This allows the model to effectively dictate its own learning pace, ensuring it is sufficiently prepared before tackling more complex examples. A sophisticated variant of self-paced learning is the uncertainty-aware pacing strategy, which utilizes an “uncertainty-aware sampling pacing function” to dynamically adapt the curriculum based on difficulty metrics derived from the model’s in-domain uncertainty.¹⁰ This ensures that the network prioritizes learning from the most certain, and thus often easier, samples first, with the curriculum dynamically adjusting as the model’s certainty evolves throughout training.¹⁰

Beyond explicit data ordering, some CL methods define an “implicit curriculum” by progressively increasing the inherent complexity of the learning task itself, rather than explicitly sorting individual data points. In dense prediction tasks like image segmentation, “Progressive Growing of Patch Size” involves gradually increasing the size of the input patches fed to the model during training. The assumption is that training on smaller foreground patches constitutes an easier optimization task. As patch sizes grow, they encompass more global context, thereby increasing task difficulty. This implicit curriculum has been shown to accelerate convergence and improve performance compared to standard training with constant patch sizes, while also significantly reducing runtime, computational costs, and CO2 emissions.¹² Another novel and simple strategy for tasks like unsupervised medical image alignment is “Curriculum by Input Blur.” This involves initiating training with purposely blurred images and then gradually transitioning to sharper, higher-fidelity images in subsequent training stages.¹³ The rationale is that blurring reduces the visibility and contrast of fine details , making initial coarse alignment easier. As the images become sharper, the task becomes more challenging, requiring the model to learn more precise alignments. This method has demonstrated superior results and an excellent accuracy-speed trade-off in medical image registration.¹³

The existence of both explicit and implicit curriculum learning methodologies highlights a strategic trade-off for researchers and developers. Explicit methods, such as weighting samples, reordering datasets, or sampling subsets based on a defined difficulty score ⁸, offer fine-grained control over the data presentation. This precision can be particularly beneficial for highly imbalanced or noisy datasets, allowing direct intervention in the learning process to address specific data pathologies.⁸ Conversely, implicit methods, by embedding the curriculum into data transformations or architectural choices, often lead to significant practical advantages in terms of resource efficiency and scalability. These methods, like progressive growing of patch size or curriculum by input blur, can reduce runtime, lower computational costs, and decrease CO2 emissions, while still improving convergence and performance.¹² The increasing adoption of implicit curricula due to their computational benefits suggests a future where hybrid approaches may become prevalent. These could combine the fine-grained control of explicit methods, for instance, for initial phases or handling specific hard cases, with the efficiency gains of implicit methods, such as for large-scale pre-training or general progression.

Applications of Curriculum Learning in 3D Medical Imaging

3.1. Enhancing 3D Medical Image Segmentation

Adapting large, pre-trained foundation models, such as the Segment Anything Model (SAM), for highly specialized medical image segmentation tasks presents unique challenges. “Curriculum prompting” is an innovative CL-inspired approach designed to address this. It systematically integrates prompts of varying granularity, following a coarse-to-fine mechanism.¹⁶ This method automates the generation of optimal prompts for SAM-based medical image segmentation, thereby eliminating the need for labor-intensive manual intervention. It provides the network with more image-specific details and clinically relevant knowledge.¹⁶ The process typically begins with coarse prompts, such as bounding box prompts, to obtain initial, rough masks. These masks are then progressively refined using finer-grained prompts, such as point prompts, and the previously generated coarse masks as additional input. This progressive integration helps to mitigate potential conflicts that can arise when simultaneously using multiple prompt types from different domains.¹⁶

Another effective implicit curriculum learning strategy specifically tailored for dense prediction tasks, including segmentation, is the “Progressive Growing of Patch Size.” This methodology involves gradually increasing the size of the input patches fed to the model during training. The underlying assumption is that training on smaller foreground patches is an inherently easier optimization problem for the model. As the patch size grows, the task difficulty incrementally increases. This approach has demonstrated significant benefits, including improved convergence and superior performance compared to traditional training with a constant patch size, all while substantially reducing runtime, computational costs.¹²

The distinct curriculum learning methodologies applied to 3D medical image segmentation, such as “curriculum prompting” and “progressive patch size,” reflect the nuanced requirements of this domain. Segmentation focuses on delineating fine spatial details within complex anatomical structures. The variations in CL implementation across different medical imaging tasks are not coincidental; they indicate that the optimal definition of “difficulty” and the most effective scheduling strategy are highly dependent on the specific nature of the task and the inherent characteristics of the 3D medical data. For instance, segmentation’s emphasis on precise spatial boundaries necessitates a tailored approach to difficulty progression that might differ from classification or registration tasks. This underscores that effective CL application in medical imaging demands a deep understanding of both machine learning principles and the specific clinical and data nuances of the task at hand. This implies that future CL research in medical imaging will increasingly move towards highly specialized, task-aware, and data-adaptive curriculum designs. It reinforces the critical need for interdisciplinary research, where medical imaging experts collaborate closely with machine learning scientists to define clinically meaningful difficulty metrics and develop bespoke curriculum strategies that truly align with the complexities of specific medical problems.

Improving 3D Medical Image Classification

Curriculum learning frameworks are particularly advantageous for multi-class classification tasks in medical imaging. This domain frequently grapples with challenges such as the difficulty and expense of acquiring large, high-quality annotated datasets, the prevalence of highly imbalanced class distributions, and the presence of noisy labels stemming from intra- or inter-expert disagreement.⁸ Curriculum learning provides a robust and unified framework specifically designed to mitigate these inherent data challenges by intelligently scheduling data presentation based on defined difficulty or uncertainty.⁸

Comprehensive CL frameworks propose to mitigate these issues by intelligently scheduling the order and pace of training samples. These frameworks typically integrate three core strategies: individually weighting training samples, reordering the entire training set, or progressively sampling subsets of data.⁸ The efficacy of these strategies hinges on a scoring function that ranks training samples. This function can be static, derived from domain-specific prior knowledge, for example, based on Cohen’s kappa for expert agreement in fracture classification or F1-scores for digit recognition. Alternatively, it can be dynamic, quantified by directly measuring the model’s uncertainty in its predictions, such as using predictive entropy via Monte-Carlo dropout.⁸ Such unified frameworks have empirically demonstrated their ability to reduce error rates in scenarios characterized by limited data, class imbalance, and label noise.⁸

Noisy labels represent a critical issue in medical datasets, capable of significantly degrading model performance.¹⁷ CUFIT (Curriculum Fine-Tuning) is a specialized CL framework specifically designed to fine-tune Vision Foundation Models (VFMs) when confronted with such noisy medical datasets for classification tasks. This method ingeniously leverages the inherent robustness of linear probing and the strong generalization capabilities of fine-tuning adapters.¹⁷ CUFIT operates by selecting “clean” samples based on an agreement criterion, where the sample’s annotation matches the module’s prediction, and its experimental results show superior robustness against high noise rates across various medical image benchmarks.¹⁷

A novel pseudo-labeling semi-supervised learning method for medical image classification introduces an “anti-curriculum strategy.” In contrast to the traditional easy-to-hard progression, this approach sorts samples by their maximum probability predictions from small to large, implying a progression from less certain or harder to more certain or easier samples. The model is then trained in a manner that aims to prevent it from producing high-value predictions for samples that are overly similar to existing labeled data. This counterintuitive strategy is designed to improve model generalization and prevent overfitting on easily recognized, potentially redundant samples.¹⁸

The introduction of the “anti-curriculum strategy” serves as a strategic counterpoint to the dominant easy-to-hard paradigm in curriculum learning. While traditional CL seeks to build foundational knowledge on easy examples, this alternative approach for semi-supervised medical image classification suggests that, for certain objectives, particularly when generating pseudo-labels, focusing on “harder” or more diverse examples early on might be beneficial. This is intended to prevent the model from memorizing trivial patterns and instead enhance true generalization.¹⁸ This challenges the universal applicability of the easy-to-hard paradigm and highlights a potential limitation: the risk of overfitting on easy, redundant samples. It suggests that “difficulty” might need to be re-evaluated not just in terms of inherent complexity, but also in terms of its “informativeness” or “diversity” relative to the model’s current knowledge state. For certain tasks, particularly those requiring robust generalization to out-of-distribution data, a non-monotonic or even inverse curriculum might be more effective. This opens new avenues for curriculum design research, moving beyond simple linear progressions. Future CL strategies might involve more complex, adaptive schedules that dynamically switch between easy-to-hard and hard-to-easy phases, or even multi-objective curricula that balance learning core concepts with exploring challenging, novel, or uncertain data points to maximize overall model robustness and generalization in complex medical environments.

Advancements in 3D Medical Image Registration and Alignment

Deformable image registration is a fundamental task in 3D medical image analysis. It involves precisely aligning multiple 3D images—which might have been acquired at different time points or from different imaging modalities—to a common coordinate system. This alignment is essential for accurate comparison, longitudinal analysis of disease progression, and multi-modal data fusion.

A particularly innovative and straightforward curriculum learning strategy for unsupervised medical image alignment involves initiating the training process with purposely blurred input images. As training progresses, the model is gradually exposed to sharper, higher-resolution images in later stages.¹³ This “curriculum by input blur” method has demonstrated superior performance compared to conventional training paradigms. Furthermore, it offers an optimal trade-off between accuracy and computational speed among various curriculum learning approaches evaluated for this task.¹³ The underlying principle leverages the fact that blurring inherently reduces the visibility and contrast of fine anatomical details.¹⁴ By starting with blurred images, the model first learns to establish coarse, robust alignments based on global structures, which is an easier task. As the images become progressively sharper, the task difficulty increases, compelling the model to learn more precise, fine-grained deformations necessary for accurate registration.

Recent advancements include the development of innovative “plug-in curriculum schedulers” that can be seamlessly integrated into existing medical image registration methods without necessitating alterations to their core network architecture. These schedulers dynamically increase task difficulty during training by incorporating sophisticated criteria for sample difficulty, assessed at both voxel and volume levels, potentially using metrics like Variance of Gradients or Gaussian blurring, and by gradually increasing the stringency of matching accuracy constraints.¹⁹

The specialized curriculum learning strategies developed for 3D medical image registration, such as “curriculum by input blur” and “plug-in curriculum schedulers,” further illustrate how CL adaptations are tailored to specific task requirements. Registration, unlike segmentation or classification, is fundamentally about learning spatial transformations and correspondences. The use of input blur, which simplifies the initial alignment task by reducing fine detail, directly addresses the challenge of finding robust global transformations before refining local ones. Similarly, plug-in schedulers that dynamically adjust difficulty based on voxel-level complexity cater to the intricate nature of deformable registration. This highlights that the design of effective CL strategies in medical imaging is deeply intertwined with the underlying mathematical and computational challenges of each specific task.

Role in Self-Supervised Learning for 3D Medical Data

Self-supervised learning (SSL) has emerged as a powerful label-efficient deep learning paradigm, particularly critical for 3D medical imaging where large-scale, expertly annotated datasets are scarce and prohibitively costly to acquire.⁹ SSL enables models to learn meaningful and highly transferable representations directly from vast amounts of unlabeled data by generating automatic supervision signals from the data itself.⁹

The development of cutting-edge SSL methods specifically adapted for 3D datasets, such as 3DINO, is a significant trend. These methods are used to pretrain general-purpose medical imaging models, for example, 3DINO-ViT, on exceptionally large, multimodal, and multi-organ datasets, such as approximately 100,000 3D scans from over 10 organs. Such pretraining allows these models to generalize effectively across various modalities and organs, even to out-of-distribution tasks.²⁰ While the provided information does not explicitly detail CL

within the SSL pre-training phase, the core principle of “starting small” and progressively increasing complexity ¹ is inherently beneficial for effective SSL strategies. For instance, pre-training on simpler pretext tasks, or using lower-resolution data before moving to more complex representations, aligns perfectly with curriculum learning principles. The existence of such robust and generalizable 3D SSL frameworks ²⁰ then provides a strong foundation that can be further optimized and specialized using CL during subsequent fine-tuning for specific downstream tasks, as exemplified by curriculum fine-tuning for classification.¹⁷

The rise of large pre-trained foundation models, such as SAM and other Vision Foundation Models ⁹, represents a major trend in AI. However, adapting these general-purpose models to specialized medical tasks, which often involve unique data characteristics like noisy labels or domain shifts, remains a significant challenge.¹⁶ Curriculum learning, through strategies such as “curriculum prompting” for SAM-based segmentation ¹⁶ and “curriculum fine-tuning” for VFMs dealing with noisy labels ¹⁷, directly addresses this adaptation hurdle. These CL methods systematically introduce complexity or handle data imperfections during the fine-tuning phase, effectively bridging the gap between general pre-training and specialized medical application. This positions CL as a critical enabler for the widespread adoption of foundation models in healthcare. It suggests that future research will focus on developing sophisticated CL strategies specifically for efficient and robust transfer learning, potentially involving multi-stage curricula that account for both general and domain-specific knowledge acquisition, thereby accelerating the deployment of advanced AI in clinical practice.

Benefits and Challenges of Curriculum Learning in 3D Medical Imaging

A fundamental challenge in applying deep learning to medical imaging is the critical dependence on large-scale, high-quality annotated datasets. Curating such datasets is exceptionally difficult and expensive due to the sheer volume of data, the limited availability of expert radiologists for annotation, the tedious and time-consuming nature of the annotation process, and the rarity of certain diseases.⁸ This “annotation bottleneck” significantly hinders model development and deployment.

Curriculum learning provides a robust and unified framework specifically designed to mitigate these inherent data challenges by intelligently scheduling data presentation based on defined difficulty or uncertainty.⁸ In scenarios with scarce training data, CL can significantly reduce error rates. By initiating training with easier subsets of data and gradually increasing the complexity by introducing more challenging examples, CL effectively maximizes the learning potential from limited annotated samples.⁸ CL frameworks are also adept at handling highly imbalanced class distributions, a common problem in medical datasets (e.g., rare disease cases). By prioritizing high-priority subsets—selected either based on prior domain knowledge or dynamic uncertainty estimation—CL ensures that the model adequately learns from under-represented classes.⁸

Noisy labels, arising from intra- or inter-expert disagreement or annotation errors, can severely degrade model performance.¹⁷ CL is highly effective in addressing this. Strategies within CL, such as weighting individual samples based on their estimated uncertainty, can reduce or even eliminate the detrimental influence of flawed labels, as uncertainty can be assessed at the individual sample level.⁸ CUFIT, for example, directly tackles noisy labels by intelligently selecting clean samples based on an agreement criterion.¹⁷

Current Limitations and Open Research Questions

Despite significant progress in dynamic difficulty metrics, defining the optimal measure of “difficulty” and designing the most effective pacing function remains a complex challenge. The pacing function is intrinsically linked to the difficulty measurer, and critical decisions regarding dataset partitioning (training, validation, testing splits), the inclusion of uncertainty, and preventing information leakage are intricate.¹⁰ While CL methods show promise, some proposed approaches may exhibit limited generalizability across a wide spectrum of network architectures, diverse tasks, and varied datasets.¹⁹ Furthermore, the performance of CL models can be highly sensitive to hyperparameter tuning, often necessitating extensive re-calibration for different datasets or specific tasks.¹⁹

Works cited

5. Redefining Radiology: A Review of Artificial Intelligence Integration in Medical Imaging, accessed on July 10, 2025, https://www.mdpi.com/2075-4418/13/17/2760
6. BIMCV-R: A Landmark Dataset for 3D CT Text-Image Retrieval – MICCAI, accessed on July 10, 2025, https://papers.miccai.org/miccai-2024/paper/1194_paper.pdf
8. (PDF) Curriculum learning for annotation-efficient medical image analysis: scheduling data with prior knowledge and uncertainty – ResearchGate, accessed on July 10, 2025, https://www.researchgate.net/publication/343390521_Curriculum_learning_for_annotation-efficient_medical_image_analysis_scheduling_data_with_prior_knowledge_and_uncertainty
9. Label-Efficient Deep Learning in Medical Image Analysis: Challenges and Future Directions, accessed on July 10, 2025, https://arxiv.org/html/2303.12484v5
10. Dynamic Curriculum Learning via In-Domain Uncertainty for Medical Image Classification | MICCAI 2023 – Accepted Papers, Reviews, Author Feedback, accessed on July 10, 2025, https://conferences.miccai.org/2023/papers/226-Paper2947.html
11. Training Strategies for Radiology Deep Learning Models in Data-limited Scenarios – PMC, accessed on July 10, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC8637222/
12. Progressive Growing of Patch Size: Resource-Efficient Curriculum Learning for Dense Prediction Tasks – MICCAI, https://papers.miccai.org/miccai-2024/paper/2008_paper.pdf
13. https://arxiv.org/abs/2102.10438
15. 3D Deep Learning on Medical Images: A Review – PMC, accessed on July 10, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC7570704/
16. Curriculum Prompting Foundation Models for Medical Image Segmentation – MICCAI, accessed on July 10, 2025, https://papers.miccai.org/miccai-2024/paper/2832_paper.pdf
17. Curriculum Fine-tuning of Vision Foundation Model for Medical Image Classification Under Label Noise – arXiv, accessed on July 10, 2025, https://arxiv.org/pdf/2412.00150
18. Semi-Supervised Learning for Medical Image Classification Based on Anti-Curriculum Learning – MDPI, accessed on July 10, 2025, https://www.mdpi.com/2227-7390/11/6/1306
19. A Plug-In Curriculum Scheduler for Improved Deformable Medical Image Registration, accessed on July 10, 2025, https://openreview.net/forum?id=a0JBoEy0af
20. [2501.11755] A generalizable 3D framework and model for self-supervised learning in medical imaging – arXiv, accessed on July 10, 2025, https://arxiv.org/abs/2501.11755