MedNet-FS: a few-shot learning framework for 3D MRI-based knee injury classification

Abstract

The development of deep learning models for 3D knee MRI analysis is critically constrained by the scarcity of large, annotated datasets. Few-shot learning (FSL) offers a promising pathway to leverage small data, but its effective application to volumetric medical imaging remains underexplored. This study introduces MedNet-FS, a 3D FSL framework that strategically integrates domain-specific pre-training on knee MRI data with a Generalized End-to-End (GE2E) loss to create a highly effective solution for data-scarce environments. Our central finding is that this targeted combination is paramount; MedNet-FS significantly outperforms models using generic pre-training or standard cross-entropy loss. On the internal MRNet dataset, our framework achieved an AUC of 0.76 for ACL tear detection using only 40 samples per class, demonstrating performance competitive with supervised learning. External validation on the KneeMRI dataset confirmed its generalizability for distinguishing clear cases (AUC of 0.62 for intact vs. fully ruptured ACLs), while also highlighting a key limitation: performance decreased for ambiguous partial tears (AUC 0.58), reflecting a known diagnostic challenge. While the current performance is below the threshold for autonomous clinical deployment, this work provides a robust proof-of-concept and a strong baseline, establishing a practical and scalable FSL framework that effectively reduces annotation dependency and charts a course for future development in data-efficient medical image analysis.

Introduction

Knee joint injuries are an important clinical issue in all age groups, with high incidence rates noted particularly in athletes, the elderly, and, increasingly, the pediatric age group^1,2. A report by Nirav Maniar et al.¹, for example, refers to the growing incidence of anterior cruciate ligament (ACL) injuries in Australian populations, citing ACL rupture as the most prevalent category of knee injury, with up to a 10.4 annual increase noted in children. Along with ACL disruption, other ligamentous pathologies like posterior cruciate ligament (PCL), medial collateral ligament (MCL), lateral collateral ligament (LCL) injuries, and meniscal tears are commonly described, adding to the knee trauma’s complex clinical picture. The knee joint’s complex anatomical composition and often blurred margins of lesions present a radiological interpretation and diagnostic challenge^3,4. Magnetic resonance imaging (MRI) is now the modality of choice in evaluating internal derangements of the knee because of its excellent soft tissue contrast and multiple plane capability. Yet, notwithstanding its diagnostic promise, the consistent and accurate interpretation of knee MR images is not straightforward, especially in the disabled and elderly populations, where pathological hints may be hard to detect⁵.

Recent progress in deep learning, specifically through the use of convolutional neural networks (CNNs), has greatly improved the capabilities of automated image interpretation systems. Such models have been shown to be effective in the classification of musculoskeletal pathology across both two-dimensional (2D) and three-dimensional (3D) MRI datasets^6–8. A key development in this field is the development of the MRNet architecture by Bien et al.⁹, which enabled the classification of anterior cruciate ligament (ACL) and meniscal tears through deep learning by combining features from sagittal, coronal, and axial imaging planes. The tri-plane approach substantially reduces the diagnostic error rate by incorporating large-scale spatial information. Similarly, the models proposed by Liu et al.^10,11 showed high accuracy in the detection of ACL and cartilage lesions, respectively, using improved deep learning architectures specifically tailored to specific knee pathologies. However, much of these studies has concentrated only on ACL tears and cartilage lesions, whereas other types of knee pathology have been comparatively neglected. A main reason underlying this discrepancy is the limited availability of properly annotated medical imaging datasets, as well as the huge amount of human effort and expertise required in compiling large-scale annotated data⁷. Consequently, current approaches often lack generalizability and scalability, highlighting the need for strong and versatile solutions for comprehensive knee injury assessment.

Public datasets have been instrumental in deep learning system development and validation for knee pathology classification. Among the first of these, the SKI10 dataset was released as part of the MICCAI 2010 Grand Challenge to aid segmentation tasks of cartilage and bone in knee MR images, specifically for osteoarthritis evaluation¹². In 2018, the MRNet dataset, published by researchers at Stanford University, emerged as a standard for training and validation of deep learning systems for ACL tears, meniscal injuries, and associated abnormalities⁹. Another notable resource is the KneeMRI dataset, released in 2017 by the Clinical Hospital Centre Rijeka in Croatia, consisting of 917 labeled MR knee scans of healthy knees, partial ACL tears, and completely ruptured ligaments¹³. These datasets have had a profound impact on the evolution of machine learning in medical imaging, facilitating standardized evaluation protocols and the creation of more efficient diagnostic models.

To address the persistent issues related to limited labeled data, recent research has explored various machine learning techniques that aim to enhance model performance under low-data regimes. These include semi-supervised learning⁷, transfer learning⁸, GANs¹⁴, FSL¹⁵, meta-learning¹⁶, and methods that address data imbalance¹⁷. Among these, FSL has been shown to have great promise in the domain of medical image analysis since it is explicitly developed to classify novel categories using only a handful of annotated examples. This is especially critical in clinical settings where the collection of large amounts of expert-labeled data is time-consuming and costly. FSL algorithms generally operate by dividing the dataset into a support set, containing a small number of labeled examples, and a query set, made up of unlabeled samples that are matched to the support set based on their embedding similarities. The success of this approach has been supported by many studies on tasks requiring accurate classification with few supervision^18–20. In the FSL setting, the number of labeled examples per class in the support set is termed as k, and a classification problem with k samples for each class are normally called a k-shot task.

Building on these foundations, we introduce MedNet-FS, a specialized few-shot learning framework for binary classification of knee injuries from 3D MRI scans. The model leverages a generalized end-to-end (GE2E) loss function for allowing the production of discriminative embeddings from 3D MR images (Fig. 1). The diagnostic performance of MedNet-FS was comprehensively assessed across several datasets, including the publicly available MRNet⁹ and KneeMRI¹³ datasets. A key novelty in this framework is its reliance on domain-specific pretraining, allowing the network to produce strong representations aligned with knee-MRI-specific features. We also performed comparative studies between models trained from general-purpose 3D pre-trained weights and fine-tuned medical imaging-specific pre-training. These experiments demonstrated that the use of custom-trained weights, especially those specialized to knee-specific MRI features, drastically enhances performance under low-data conditions. Our results highlight the utility of task-aligned pretraining strategies and demonstrate the effectiveness of MedNet-FS as an efficient and scalable diagnostic aid for knee pathology detection. The novelty of MedNet-FS lies not in any single component in isolation, but in the specific and deliberate combination of a 3D ResNet101 backbone adapted for volumetric MRI inputs, domain-specific pre-training on knee MRI data, and the GE2E loss function – originally developed for speaker verification and here repurposed for 3D musculoskeletal few-shot learning. This combination uniquely addresses the annotation bottleneck that limits clinical deployment of deep learning in musculoskeletal imaging by achieving supervised-level performance with minimal labeled data.

Fig. 1.

Illustration of the supervised training pipeline using the MedNet framework (based on ResNet101) for a specific knee injury classification task using 3D MR images. The figure also depicts the application of few-shot learning by leveraging the pre-trained weights obtained from this task to enable classification of other, previously unseen knee injury categories.

Materials and methods

Data selection

To create and validate the proposed MedNet-FS framework, we utilized two publicly accessible knee MRI datasets: MRNet⁹ and KneeMRI¹³. The MRNet dataset, acquired by Stanford University Medical Center, was the primary source for training and internal validation of our model. It comprises 1,370 MRI scans derived from 1,199 patients and three orthogonal imaging planes–sagittal, coronal, and axial–per examination at a 256 × 256 pixel resolution. They were labeled according to whether they had or did not have anterior cruciate ligament (ACL) tears, meniscal tears, and other pathologies. The dataset includes 1.5T and 3.0T scans performed on GE Healthcare MRI systems, and its diagnostic and demographic composition is presented in Table 1. We employed the KneeMRI dataset¹³ for external validation, which was obtained between 2006 and 2014 at the Clinical Hospital Centre Rijeka in Croatia. It has 917 labeled MRI scans for ACL conditions and three clinically relevant classes: healthy, partially torn, and completely ruptured. Diagnoses were validated using a double-blind protocol by employing arthroscopically confirmed findings. All images were obtained on a 1.5T Siemens Avanto magnetic resonance imaging (MRI) scanner. Demographic data such as age and sex are not reported for this dataset, but its diagnostic statistics are reported in Table 1.

Table 1.

Comparison of MRNet and KneeMRI datasets.

Attribute	MRNet9	KneeMRI13
Purpose	Training/validation	External evaluation
Total scans	1250	917
Total patients	1199	N/A
Female patients (%)	530 (44.20%)	N/A
Mean age (SD)	38.11 (16.90)	N/A
Mean BMI (SD)	Not Provided	Not Provided
Institution	Stanford Univ. Medical Center, USA	Clinical Hospital Centre Rijeka, Croatia
MRI scanner	GE Healthcare System	Siemens Avanto (Healthineers)
Magnetic strength	1.5T and 3.0T	1.5T
Class labels	3 Categories	3 Categories
Class distribution
ACL tear	262 (20.96%)	N/A
Meniscal tear	449 (35.92%)	N/A
Abnormal	1008 (80.64%)	N/A
Intact ACL	N/A	690 (75.25%)
ACL partial tear	N/A	172 (18.76%)
Complete ACL rupture	N/A	55 (5.99%)

Data usage

The training protocol included a phased training using the MRNet dataset⁹, as illustrated in Fig. 2, followed by its extension to a few-shot learning paradigm for the evaluation of new conditions (see Fig. 1). We prepared two sets of pre-trained weights: one from a binary classification task between ACL tear and intact ACL (referred to as Weight 1), and the other from the classification of meniscus tear versus non-meniscus injury (named as Weight 2). The two pre-trained weights were achieved under separate training protocols to maintain model specificity and avoid task interference. Under the few-shot learning paradigm, we adopted a binary (2-way) classification scenario with a k-shot setting, where k – following standard few-shot learning nomenclature as established in the seminal works of Vinyals et al. (2016) and Snell et al. (2017), and distinct from the k used in k-fold cross-validation – denotes the number of labeled support examples per class, set specifically to . To avoid ambiguity throughout this manuscript, k is reserved exclusively for this FSL support-set size, while the cross-validation protocol is always referred to in full as ’5-fold cross-validation’ without abbreviation. The training and testing datasets were purposely kept patient-disjoint to avoid data leakage. The MedNet-FS model was initialized using the said weights and tested using both internal (MRNet) and external (KneeMRI) test datasets. Despite the ACL-centric nature of the KneeMRI dataset, we used Weight 2 during the tests to investigate the transferability of feature between tasks and to prevent overfitting of specific tasks.

Fig. 2.

Overview of the experimental workflow illustrating the process of generating distinct pre-trained weights and applying them across training, internal validation, and external testing stages. The training and testing splits were determined based on the original dataset-specific class partitions. The separation between training and testing sets adheres to the original dataset’s predefined classification scheme to ensure unbiased evaluation.

We also assessed two additional pre-training baselines: (1) a general pre-trained model trained on volumetric video data²¹, and (2) a medical image model pre-trained on non-knee MRI scans²². It is important to note that Table 2 describes the fixed dataset partitioning used exclusively for primary model development and benchmarking (Sections 3.1–3.5), and was maintained to ensure direct comparability with prior published results on the MRNet dataset. The 5-fold cross-validation reported in Section 3.6 was conducted as a completely separate and independent robustness analysis on the full 1,250 scans, where each fold used approximately 1,000 scans for training and 250 for testing in a patient-disjoint manner. These two protocols are not in conflict; they serve distinct and explicitly defined experimental purposes. These allowed us to study the influence of pre-training domain specificity on few-shot performance. The exact partitioning of datasets for pre-training, validation, and testing, are summarized in Table 2. For the MRNet dataset, a total of 1,250 scans were used after excluding low-quality or incomplete studies from the original 1,370 scans. These were partitioned into training (875 scans), validation (250 scans), and testing (125 scans) sets, ensuring patient-disjoint splits across partitions. The KneeMRI dataset (917 scans) was used exclusively for external testing in few-shot scenarios, with no further partitioning. For few-shot experiments, support samples were randomly drawn from the training set, and query samples were drawn from the test set, with the process repeated over five random seeds to ensure robustness.

Table 2.

Dataset partitions for training, validation, and testing.

Dataset	Split	Total scans	ACL Tear	Meniscal tear	Other/healthy
MRNet	Training	875	183	314	378
	Validation	250	53	90	107
	Testing	125	26	45	54
Total used		1250	262	449	539
KneeMRI	External test	917	690 (intact)	172 (partial)	55 (full rupture)

MedNet-FS model construction

We proposed MedNet-FS, a few-shot learning framework based on ResNet101²³ combined with the Generalized End-to-End (GE2E) loss²⁴. We based our framework on well-established 3D CNN architectures rather than transformer-based models, motivated by the proven efficacy of CNNs in medical image analysis and the current scarcity of large-scale 3D datasets needed for effective transformer pre-training in this domain. The standard classification head was removed and replaced with a global average pooling layer, which produces a fixed-dimensional embedding vector of size (the channel dimension of the final convolutional layer in ResNet101). Class prototypes were computed as the mean of these support embeddings. Class prototypes were computed from these embeddings as a mean of support embeddings. Classification was done based on comparing the Euclidean distance of query samples to class prototypes. The GE2E loss was selected over other metric learning objectives like Prototypical or Triplet loss for its architectural efficiency and superior performance in creating well-separated embedding clusters. Unlike Prototypical loss which computes gradients from a single prototype per class, or Triplet loss which relies on careful triplet mining, GE2E considers the similarity of an embedding to all class centroids in the support set within a single update step. This provides a more global and stable optimization signal, which we found to be crucial for effective learning in our low-data, 3D medical imaging regime. The GE2E loss was used to train the network to attract same-class embeddings together and push different-class embeddings apart. Let be the embedding of the sample of the class. Let denote the similarity between this embedding and centroid of class k, where we use cosine similarity. The loss is given by:

1

This loss design allowed for better clustering of features and eliminated the need for a linear classifier, which is especially beneficial under few-shot regimes. The necessity of the full ResNet101 depth is grounded in the complexity of 3D volumetric MRI data. The initial convolutional layer (Conv1) captures only low-level features such as edges and basic textures, which are insufficient for distinguishing subtle pathological differences between injured and intact knee structures. The subsequent bottleneck blocks progressively learn higher-order anatomical representations – Layer 1 captures local tissue patterns, Layers 2 and 3 encode increasingly abstract structural relationships across the knee joint, and Layer 4 produces the high-dimensional semantic embeddings necessary for effective prototype-based few-shot classification. Removing any of these blocks would collapse the model’s ability to generate discriminative embeddings in the GE2E loss framework, particularly under low-data conditions where the embedding space must be highly structured. For benchmarking our model, we compared its performance against well-known 3D CNN architectures such as DenseNet²⁵, VGG²⁶, and AlexNet²⁷. The proposed MedNet-FS framework is built upon a 3D ResNet-101 backbone, adapted for volumetric MRI inputs of size . A detailed layer-wise breakdown of the architecture–including kernel sizes, output dimensions, parameter counts, and FLOPs–is provided in Table 3. The model replaces the standard classification head with a global average pooling layer, producing a 2048-dimensional embedding vector used for prototype-based few-shot classification. The total parameter count is approximately 85.2M, with an estimated computational cost of 126.12 GFLOPs per forward pass and an average inference time of 45 ms per volume on an NVIDIA RTX A4000 GPU. Hyperparameter and training configuration details are presented in Table 4. The hyperparameters reported in Table 4 were selected through a systematic grid search. Learning rates from ∈ 0.001, 0.0005, 0.0001, 0.00005 and optimizers ∈ Adam, SGD, AdamW were evaluated. Higher learning rates such as 0.001 resulted in unstable convergence of the GE2E loss, while SGD demonstrated poor embedding space stability under metric learning objectives. AdamW produced only marginal differences compared to Adam. A learning rate of 0.0001 with the Adam optimizer was therefore selected as it consistently produced the highest AUC and most stable embedding convergence across both classification tasks.

Table 3.

Layer-wise architecture and computational details of MedNet-FS, including parameter counts, FLOPs, multiply-accumulate operations (MACs), inference time, and GPU memory usage per forward pass.

Layer/Block	Type	Kernel/stride	Output shape	Parameters	FLOPs (G)
Input	3D MRI volume	–	1×42×256×256	0	0.00
Conv1	3D Conv + BN + ReLU	7×7×7,/2	64×21×128×128	9,472	0.32
MaxPool	3D Max Pooling	3×3×3,/2	64×11×64×64	0	0.02
Layer1 (×3)	Bottleneck block	[1,3,1], 64→256	256×11×64×64	6.94M	10.12
Layer2 (×4)	Bottleneck block	[1,3,1], 128→512	512×6×32×32	14.75M	21.85
Layer3 (×23)	Bottleneck block	[1,3,1], 256→1024	1024×6×32×32	52.42M	77.52
Layer4 (×3)	Bottleneck block	[1,3,1], 512→2048	2048×3×16×16	11.01M	16.28
Global Avg pool	3D adaptive pool	–	2048×1×1×1	0	0.01
Embedding	Flatten	–	2048	0	0.00
Total				85.20M	126.12

Inference time: 45 ms per volume (NVIDIA RTX A4000 GPU). Multiply-Accumulate Operations (MACs): 63.06 G. GPU Memory Usage:∼6.2 GB per forward pass.

Table 4.

Hyperparameters of MedNet-FS and other benchmarking models.

Model	No. of params	Optimizers	Epoch	Learning rate	Loss function
MedNet_FS	85,201,729	Adam	100	0.0001	GE2E
MedNet	85,201,729	Adam	100	0.0001	Cross entropy
VGG16	188,072,514	Adam	100	0.0001	Cross entropy
AlexNet	175,148,866	Adam	100	0.0001	Cross entropy
DenseNet	12,267,624	Adam	100	0.0001	Cross entropy

Performance metrics

Model evaluation was based on four key metrics: sensitivity (SEN), specificity (SPE), accuracy (ACC), and area under the receiver operating characteristic curve (ROC-AUC). These metrics were computed using the MRNet validation set to assess diagnostic performance comprehensively and ensure comparability across methods.

Statistical analysis

Statistical comparisons were conducted using GraphPad Prism v10.3.1 (GraphPad Software, San Diego, CA, USA)²⁸. Differences between models were evaluated via multiple independent t-tests with 95% confidence intervals. Statistical significance was reported at thresholds of p < 0.05 (*), p < 0.01 (**), p < 0.001 (***), and p < 0.0001 (****).

Results

Patient characteristics

For this problem, internal training and validation were conducted on the MRNet dataset, consisting of 1,250 knee MRI scans gathered from 1,199 unique patients. Every patient had MR imaging performed in three planes of anatomy–sagittal, coronal, and axial–with all images normalized to a 256 × 256 pixel resolution. As reported in Table 1, the mean age of the patient population was 38.11 years with a standard deviation of ±16.90 years, and 44.20 of subjects were female. While body mass index (BMI) information was not reported, the demographic range was adequately extensive to include a representative subset of the adult population. In keeping with the structure of our planned classification experiments, the MRNet dataset was divided into two distinct binary tasks: the first aimed at the differentiation of anterior cruciate ligament (ACL) tears from non-ACL diagnoses, and the second the detection of meniscal tears versus non-meniscal presentations. By splitting binary tasks in this way, we were able to create two distinct sets of pre-trained weights from the MRNet data–each specialized to a specific type of injury–which would subsequently be used to facilitate few-shot learning assessments.

To examine the generalizability of the proposed method in a clinically distinct setting from the environment used for training, we employed the KneeMRI dataset for external validation. The dataset comprising 917 MRI volumes was acquired at the Clinical Hospital Centre Rijeka, Croatia, between 2006 and 2014, and all the images were acquired using a Siemens Avanto 1.5T scanner. Notably, the dataset includes diagnostic labels related to ACL integrity, divided into three distinct classes: intact ACL, partially torn ACL, and fully ruptured ACL. The development of these diagnostic labels was accomplished via a rigorous double-blind labeling process, where arthroscopically confirmed results were used as the reference standard for accuracy, thereby ensuring high reliability in annotation and clinical relevance. Conveniently, this dataset differs from MRNet not only in the imaging equipment type used but also in its classification scheme, which follows a three-class system rather than binary labels. This variability–both in terms of the type of scanner as well as the diagnostic labeling protocol–allowed us to examine to what degree the MedNet-FS framework can generalize well across domains with considerable distributional differences. Further details on the MRI protocols can be found in Table 5. This combination of internal and external datasets formed a solid foundation for the evaluation of both the performance and transferability of the proposed few-shot classification model.

Table 5.

MRI scanner parameters utilized in the MRNet and KneeMRI datasets.

Dataset	Magnet strength	Imaging protocol	Slice thickness (mm)	Slice gap (mm)	Acquisition matrix	Reconstructed matrix	Bit depth
MRNet	3T	Sag PD	2.5	0	384×224×42	512×512×42	16
	3T	Sag T2 FS	2.5	0	384×192×42	512×512×42	16
	1.5T	Sag PD	3.0	1.0	512×192×24	512×512×24	16
KneeMRI	1.5T	Sag T2 FS	3.5	0.5	448×192×24	512×512×24	16
	1.5T	Sag PD	3.0	3.6	320×320×32	320×320×32	12

Exploration of different pre-trained weights and CNN architectures in few-shot classification of knee injuries

To understand the impact of different pre-training regimes on the performance of few-shot classification, we systematically investigated the behavior of MedNet-FS initialized with different pre-trained weights. We specifically considered weights derived from the MRNet dataset, including both anterior cruciate ligament (ACL) injury-specific and meniscus-specific ones, as well as weights derived from general medical imaging models and volumetric video datasets. The diagnostic performance of these setups for both ACL tear and meniscus tear classification tasks is illustrated in Fig. 3a–d. Interestingly, when initialized with Pre-trained Weight 2, trained on the task of classifying meniscus tear vs. non-meniscus tear cases, the MedNet model exhibited considerably better performance on the classification of ACL injuries. In the ACL tear classification setting, MedNet yielded an accuracy (ACC) of 0.68 and an area under the curve (AUC) of 0.72, as indicated in Fig. 3a–c. For the meniscus classification task, the same pre-trained weight led to an ACC of 0.65 and an AUC of 0.68, illustrated in Fig. 3b,d. These results highlight a key observation: even when the source pre-training task is different, if it is derived from knee MRI data, the representations learned can be effectively transferred to novel but related tasks in a few-shot setting. This level of performance is especially noteworthy given the limited labeled support available in typical few-shot settings, thereby further supporting the effectiveness of domain-specific pre-training.

Fig. 3.

Comparative performance of different pre-trained weights and CNN architectures for few-shot classification of ACL and meniscal tears on the MRNet dataset. (a, b) show accuracy, sensitivity, and specificity using various pre-trained weights with MedNet; (c, d) display corresponding ROC-AUC curves. (e, f) present classification performance of different CNNs (MedNet, VGG16, DenseNet, AlexNet); (g, h) show their ROC-AUC curves. p < 0.001; ns: not significant.

Conversely, the pre-training of MedNet with pre-trained weights derived from general-purpose medical imaging models–like Med3D⁹–or from non-medical volumetric datasets like Kinetics led to a drastic fall in its few-shot classification performance. In ACL tear classification, the model only managed 0.53 ACC and an AUC of 0.50 (Fig. 3a,c), indicating close to random predictions. The performance for the meniscus tear classification was equally disappointing, with an ACC of 0.48 and an AUC of 0.42 (Fig. 3b,d). The findings provide strong evidence that the effectiveness of few-shot classification in the medical imaging domain is heavily reliant on the level of alignment between the pre-training task and the target task. Generic embeddings, while useful for coarse-level feature extraction, lack the desired domain specificity to recognize subtle anatomical differences related to knee injuries. This highlights the need to create task-specific pre-trained models in order to maximize the benefits of few-shot learning in clinically meaningful applications.

We further expanded our study by comparing the performance of various convolutional neural network (CNN) backbones in the same training and testing conditions. MedNet was benchmarked against VGG16, DenseNet, and AlexNet, all of which were tested for both ACL and meniscus tear classification tasks. The performance metrics–sensitivity (SEN), specificity (SPE), and classification stability–are shown in Fig. 3e–h. In ACL tear classification, MedNet reached a balanced sensitivity of 0.69 and specificity of 0.68 (Fig. 3h), significantly outperforming other architectures. By comparison, VGG16 had a high sensitivity of 0.94 but a low specificity of 0.47, indicating a high false-positive rate. Likewise, AlexNet attained a sensitivity of 0.91 but lagged behind in specificity, with a score of only 0.39, reflecting severe misclassification of negative samples. DenseNet provided modest balance (SEN: 0.74, SPE: 0.51) but still performed worse compared to MedNet, validating the fact that MedNet’s architecture is better optimized for building strong embeddings appropriate for few-shot classification tasks.

A similar trend was observed in meniscus tear classification. MedNet manifested a sensitivity of 0.71 and specificity of 0.60, reflecting the best trade-off between true positive and true negative rates among all the backbones tested (Fig. 3f). VGG16 manifested more balanced values (SEN: 0.62, SPE: 0.62), though its performance continued to be inferior. AlexNet had higher variability with a SEN of 0.65 and SPE of 0.60, while DenseNet had a slightly better balance but still lagged behind MedNet in both metrics. The persistent better performance of MedNet in both classification tasks reaffirms the hypothesis that its architectural design and optimized embedding generation process make it highly capable of few-shot medical image analysis. The experimental conditions for all the models were kept constant to have a fair comparison, with training conducted using the Adam optimizer for 100 epochs with a learning rate of 0.0001. The number of parameters, loss functions employed, and other hyperparameter settings can be found in Table 4. Interestingly, whereas MedNet-FS employed the GE2E loss function to promote intra-class clustering and inter-class separability in the embedding space, all other models were trained using conventional cross-entropy loss. This further confirms the superiority of the GE2E loss formulation when used in few-shot learning settings.

Performance of the MedNet-FS model with GE2E loss

The GE2E loss function was instrumental in improving the discriminative power of the MedNet-FS model for few-shot classification of knee injuries. In contrast to conventional loss formulations like cross-entropy that mainly optimize model outputs by reducing prediction error against ground truth labels, GE2E loss directly acts in the embedding space. Figure 4 conceptually compares the two loss functions. As can be seen from Fig. 4a, the cross-entropy loss updates embedding vectors by punishing misclassified samples but does not have the ability to explicitly optimize intra-class compactness and inter-class separation. Thus, while it is effective in supervised classification tasks, its limitations become more apparent under low-data regimes like few-shot learning. Conversely, the GE2E loss function, illustrated in Fig. 4b, is explicitly tailored to generate embeddings that are compactly clustered within classes and well-separated across different classes. By jointly minimizing intra-class distances and maximizing inter-class distances, GE2E regularizes the embedding space to be more structured and robust, making it ideal for prototype-based classification. This is especially useful in few-shot learning, where the model has to rely on a small number of support samples to create meaningful decision boundaries. This design advantage is also brought to light by the comparison of MedNet and MedNet-FS, utilizing the GE2E loss. On the ACL tear classification task, MedNet-FS yielded a statistically significant boost in sensitivity (p < 0.05) and an AUC of 0.76, outperforming the baseline MedNet model, which achieved an AUC of 0.72 (Fig. 4c,d). Similarly, on the meniscus tear classification task, MedNet-FS not only boosted sensitivity and specificity (p < 0.05) but also exhibited a marginal gain in AUC, from 0.68 to 0.69 (Fig. 4e,f).

Fig. 4.

Impact of GE2E loss on MedNet-FS performance. (a, b) Visual comparison of Cross-Entropy vs. GE2E loss mechanisms. (c, d) ACL tear classification metrics and ROC-AUC curves comparing MedNet and MedNet-FS. (e, f) Meniscus tear classification performance and ROC-AUC curves. *p < 0.05; ns: not significant.

These gains confirm the hypothesis that embedding-level loss functions, e.g., GE2E, provide significant benefits under few-shot learning conditions. They also reaffirm the value of intra-class and inter-class embedding geometry as an essential aspect of model training. The performance boost obtained using GE2E loss highlights its value in developing generalizable, high-performing models for medical image classification, especially in clinically data-scarce environments.

Efficacy of few-shot learning method over supervised learning on small samples

The support set sizes (k = 2, 20, 40) were chosen to represent a spectrum of few-shot scenarios. A 2-shot task represents an extremely low-data regime, testing the model’s ability to learn from minimal information. k=20 and k=40 were selected to demonstrate the performance scaling as a modest amount of annotated data becomes available, with k=40 serving as a benchmark to show that the model can approach the performance of a supervised model, thereby validating its data efficiency. To assess the learning efficiency and practical benefits of the few-shot learning paradigm, we compared the MedNet-FS model with a fully supervised MedNet model under the same network architectures and training conditions. The comparison was carried out under different support set sizes, namely at k = 2, 20, and 40, where k represents the number of labeled support examples per class. As presented in Table 6, the supervised MedNet-FS model trained on k = 40 samples (20 positive, 20 negative) recorded an accuracy of 0.67 and an AUC of 0.75 for ACL tear classification. The few-shot model, which was initialized using only k = 2 samples, started underperforming, achieving an accuracy of 0.43 and an AUC of 0.31. However, its performance drastically improved with the increased number of support samples. At k = 20, the few-shot model attained 0.68 accuracy and an AUC of 0.70, and at k = 40, it achieved an accuracy of 0.72 and an AUC of 0.76, thereby outperforming the supervised counterpart. An identical performance trend was witnessed in the meniscal tear classification task. The supervised model recorded 0.67 accuracy and 0.64 AUC. The few-shot model, however, started with a meager accuracy of only 0.39 and a very low AUC of 0.25 at k = 2. With more labeled examples, however, its performance rapidly closed in on that of the supervised model–reaching 0.73 accuracy and 0.76 AUC at k = 20 and 0.69 accuracy and 0.70 AUC at k = 40. These results, presented in Table 6, strongly suggest that few-shot learning is an efficient and scalable alternative to supervised training in medical imaging applications, particularly where data annotation is costly or infeasible. As k rises, the model not only closes the performance gap but also performs at par or even better than fully supervised models trained under identical input conditions. To further benchmark our approach within the broader few-shot learning landscape, we compared MedNet-FS against two established meta-learning baselines: Prototypical Networks (ProtoNet) and Model-Agnostic Meta-Learning (MAML). As shown in Table 7, MedNet-FS performs competitively across all k-shot settings. At k=40, our framework achieved an AUC of 0.76, slightly surpassing ProtoNet (0.73) and MAML (0.74), while maintaining balanced sensitivity and specificity. This demonstrates that MedNet-FS–through its domain-specific pre-training and GE2E loss–provides a strong and practical alternative to more complex meta-learning pipelines for 3D MRI-based few-shot classification.

Table 6.

Performance comparison between traditional supervised learning and MedNet-FS under different few-shot settings (k = 2, 20, 40) for detecting ACL and meniscus injuries. Values in parentheses represent the 95% confidence interval.

Approach	Diagnosis target	Accuracy (ACC)	Sensitivity (SEN)	Specificity (SPE)	AUC
Supervised training	ACL tear	0.67 (0.64–0.70.64.70)	0.48 (0.44–0.52.44.52)	0.79 (0.76–0.82.76.82)	0.75 (0.72–0.78.72.78)
	Meniscus tear	0.67 (0.64–0.70.64.70)	0.32 (0.28–0.36.28.36)	0.93 (0.91–0.95.91.95)	0.64 (0.61–0.67.61.67)
MedNet-FS (k=2)	ACL tear	0.43 (0.39–0.47.39.47)	0.22 (0.18–0.26.18.26)	0.57 (0.52–0.62.52.62)	0.31 (0.27–0.35.27.35)
	Meniscus tear	0.39 (0.35–0.43.35.43)	0.13 (0.10–0.16.10.16)	0.61 (0.56–0.66.56.66)	0.25 (0.21–0.29.21.29)
MedNet-FS (k=20)	ACL tear	0.68 (0.65–0.71.65.71)	0.65 (0.61–0.69.61.69)	0.72 (0.68–0.76.68.76)	0.70 (0.66–0.73.66.73)
	Meniscus tear	0.73 (0.70–0.76.70.76)	0.72 (0.68–0.76.68.76)	0.74 (0.70–0.78.70.78)	0.76 (0.73–0.79.73.79)
MedNet-FS (k=40)	ACL tear	0.72 (0.69–0.75.69.75)	0.73 (0.69–0.77.69.77)	0.72 (0.68–0.76.68.76)	0.76 (0.73–0.79.73.79)
	Meniscus tear	0.69 (0.66–0.72.66.72)	0.67 (0.63–0.71.63.71)	0.72 (0.68–0.76.68.76)	0.70 (0.67–0.73.67.73)

Table 7.

Few-shot classification performance comparison between MedNet-FS and meta-learning baselines on the MRNet dataset (ACL tear task). Values in parentheses represent the 95% confidence interval.

Method	k	Accuracy (ACC)	Sensitivity (SEN)	Specificity (SPE)	AUC
ProtoNet	2	0.38 (0.34–0.42)	0.20 (0.16–0.24)	0.52 (0.47–0.57)	0.28 (0.24–0.32)
MAML	2	0.40 (0.36–0.44)	0.25 (0.21–0.29)	0.50 (0.45–0.55)	0.30 (0.26–0.34)
MedNet-FS (Ours)	2	0.43 (0.39–0.47)	0.22 (0.18–0.26)	0.57 (0.52–0.62)	0.31 (0.27–0.35)
ProtoNet	20	0.65 (0.61–0.69)	0.60 (0.56–0.64)	0.68 (0.64–0.72)	0.67 (0.63–0.71)
MAML	20	0.66 (0.62–0.70)	0.62 (0.58–0.66)	0.70 (0.66–0.74)	0.68 (0.64–0.72)
MedNet-FS (Ours)	20	0.68 (0.65–0.71)	0.65 (0.61–0.69)	0.72 (0.68–0.76)	0.70 (0.66–0.73)
ProtoNet	40	0.70 (0.67–0.73)	0.68 (0.64–0.72)	0.71 (0.67–0.75)	0.73 (0.70–0.76)
MAML	40	0.71 (0.68–0.74)	0.70 (0.66–0.74)	0.72 (0.68–0.76)	0.74 (0.71–0.77)
MedNet-FS (Ours)	40	0.72 (0.69–0.75)	0.73 (0.69–0.77)	0.72 (0.68–0.76)	0.76 (0.73–0.79)

Assessment of the MedNet-FS model on external testing datasets

To assess the generalization power of MedNet-FS, we conducted external validation on the KneeMRI dataset, focusing on the binary classification task of ACL tear vs. intact ACL. To this end, the MedNet-FS model was pre-trained with meniscus tear classification on the MRNet dataset, allowing cross-task and cross-domain transferability to be evaluated. External test settings also involved manipulating the size of support samples (k = 2, 20, 40) to replicate few-shot learning conditions. The initial experiment included all types of ACL injuries–intact, partially torn, and fully ruptured–as indicated in Table 8 under the category labeled “Including partially torn cases.” At k = 2, the model achieved an accuracy of 0.42, sensitivity of 0.67, and specificity of 0.33. Increasing k to 20 resulted in improved accuracy of 0.63, although sensitivity dropped to 0.36, with specificity rising to 0.71. At k = 40, accuracy was registered at 0.61, sensitivity at 0.47, and specificity at 0.65. AUC values showed a steady increase from 0.53 at k = 2 to 0.58 at k = 40. These findings, while indicative of performance improvement with larger support sizes, also reflect inherent challenges in classifying partially torn ACLs, where overlapping image features may weaken model confidence. The corresponding ROC-AUC curve and confusion matrix for this study are shown in Fig. 5a,b, respectively. Granting the diagnostic ambiguity that comes with partially torn anterior cruciate ligaments (ACLs), a secondary task was conducted where such cases were excluded. The filtered task solely considered intact ACLs versus completely ruptured ACLs and therefore dealt with two well-defined clinical states. This filtered task, reported in Table 8 under the “Excluding partially torn cases” label, yielded considerably better results. At k = 2, sensitivity was quite high (0.85) while specificity was low (0.15), implying the model was capable of identifying positive cases but struggled with false positives. On the other hand, at k = 20, the model’s performance had better balance, with an accuracy of 0.71, a sensitivity of 0.42, specificity of 0.73, and area under the curve (AUC) of 0.60. The most pronounced results were seen at k = 40, where the model achieved an AUC of 0.62 in conjunction with an accuracy of 0.50, a sensitivity of 0.58, and specificity of 0.50. These values convey that the model exhibits greater discriminative capabilities when dealing with well-separated classes, further supporting the value of excluding diagnostically ambiguous classes. This filtration-driven improvement in class separation and trend in performance is seen in Fig. 5c (ROC-AUC) and Fig. 5d (confusion matrix).

Table 8.

Binary classification results of ACL injury on KneeMRI dataset using MedNet-FS for different values of k, comparing performance when partially torn cases are included or excluded. Values in parentheses represent the 95% confidence interval.

Condition	Sample size (k)	Accuracy (ACC)	Sensitivity (SEN)	Specificity (SPE)	AUC
With partial tears	k = 2	0.42 (0.38–0.46.38.46)	0.67 (0.61–0.73.61.73)	0.33 (0.28–0.38.28.38)	0.53 (0.48–0.58.48.58)
	k = 20	0.63 (0.59–0.67.59.67)	0.36 (0.31–0.41.31.41)	0.71 (0.66–0.76.66.76)	0.55 (0.50–0.60.50.60)
	k = 40	0.61 (0.57–0.65.57.65)	0.47 (0.42–0.52.42.52)	0.65 (0.60–0.70.60.70)	0.58 (0.53–0.63.53.63)
Without partial tears	k = 2	0.20 (0.16–0.24.16.24)	0.85 (0.79–0.91.79.91)	0.15 (0.11–0.19.11.19)	0.46 (0.40–0.52.40.52)
	k = 20	0.71 (0.67–0.75.67.75)	0.42 (0.36–0.48.36.48)	0.73 (0.68–0.78.68.78)	0.60 (0.55–0.65.55.65)
	k = 40	0.50 (0.45–0.55.45.55)	0.58 (0.51–0.65.51.65)	0.50 (0.44–0.56.44.56)	0.62 (0.56–0.68.56.68)

Fig. 5.

External validation of MedNet-FS on the KneeMRI dataset for ACL tear classification. (a, b) ROC-AUC curve and confusion matrix with partially torn cases included. (c, d) ROC-AUC curve and confusion matrix with partially torn cases excluded.

The performance on the external KneeMRI dataset is consistent with expectations. The drop in metrics compared to the internal validation is a known effect of domain shift, caused by differences in MRI scanners and protocols between the datasets. Furthermore, the classification of ’partial tears’ is inherently challenging, even for clinicians, which explains the model’s difficulty and makes the results a realistic reflection of a clinical setting.

Robustness analysis with cross-validation and SMOTE

To independently confirm the robustness of our findings, we performed 5-fold cross-validation separately from the fixed splits described in Table 2, partitioning all 1,250 MRNet scans into patient-disjoint folds. This analysis was not intended to replace primary results but to validate their consistency across different data partitions. Cross-validation was conducted in a patient-disjoint manner, ensuring no data leakage across folds. Class imbalance is a well-documented challenge in medical imaging datasets, particularly where rare pathologies are under-represented. In the MRNet dataset, ACL tears constitute approximately 21% of scans, while the KneeMRI dataset contains only 5.99% fully ruptured cases. To address this, SMOTE was applied in the embedding space rather than at the image level, generating synthetic support embeddings for minority classes by interpolating between existing minority-class feature vectors. This approach balances class representation without requiring additional image acquisition or augmentation of raw volumetric data — which would introduce reconstruction artefacts in 3D MRI. Table 9 summarizes the performance of MedNet-FS under the original single-split setup, 5-fold CV, and CV combined with SMOTE, evaluated at k = 40 shots per class. SMOTE yielded consistent improvements across all metrics. For ACL tear detection, AUC improved from 0.75 ± 0.02 to 0.77 ± 0.02, with sensitivity rising from 0.72 ± 0.04 to 0.75 ± 0.03 — a particularly clinically meaningful gain as sensitivity directly reflects the model’s ability to correctly identify true injury cases. For meniscus tear detection, AUC improved from 0.69 ± 0.03 to 0.71 ± 0.02. Notably, for the external KneeMRI dataset, sensitivity for partial ACL tears increased from 0.47 to 0.55 under balanced few-shot settings with SMOTE, suggesting that the technique partially compensates for the severe class imbalance present in that dataset. These findings confirm that embedding-space SMOTE is a practical and effective strategy for addressing class imbalance in few-shot medical imaging settings, where image-level augmentation is computationally prohibitive and the support set is too small to benefit from traditional oversampling at the input level. Additionally, to contextualize our framework within the current few-shot learning landscape, we compared MedNet-FS against recent state-of-the-art methods. As shown in Table 10, our approach achieves competitive results with contemporary FSL models, demonstrating its relevance and effectiveness without requiring complex episodic training pipelines.

Table 9.

Robustness evaluation with 5-Fold cross-validation and SMOTE on MRNet (k=40).

Condition	AUC (mean ± SD)	Accuracy (mean ± SD)	Sensitivity (mean ± SD)	Specificity (mean ± SD)
ACL tear (original)	0.76	0.72	0.73	0.72
ACL tear (5-Fold CV)	0.75 ± 0.02	0.71 ± 0.03	0.72 ± 0.04	0.71 ± 0.03
ACL tear (CV + SMOTE)	0.77 ± 0.02	0.73 ± 0.02	0.75 ± 0.03	0.73 ± 0.03
Meniscus tear (original)	0.70	0.69	0.67	0.72
Meniscus tear (5-Fold CV)	0.69 ± 0.03	0.68 ± 0.03	0.66 ± 0.04	0.71 ± 0.04
Meniscus tear (CV + SMOTE)	0.71 ± 0.02	0.70 ± 0.02	0.69 ± 0.03	0.73 ± 0.03

Table 10.

Comparison of MedNet-FS with state-of-the-art few-shot learning methods at k=40. Method names and their corresponding references are listed separately to clearly distinguish methodological labels from citations.

Method	Reference	AUC	Accuracy	Sensitivity	Specificity
Prototypical networks	Snell et al.36	0.73	0.70	0.68	0.71
Matching networks	Vinyals et al.37	0.72	0.69	0.67	0.70
Meta-baseline	Chen et al.38	0.75	0.72	0.71	0.73
FEAT	Ye et al.39	0.76	0.73	0.72	0.74
DeepEMD	Zhang et al.40	0.77	0.74	0.73	0.75
MedNet-FS (Ours)	–	0.76	0.72	0.73	0.72

Discussions

This study presents MedNet-FS, a 3D FSL framework designed to address the urgent clinical need for accurate, efficient, and scalable classification of knee joint injuries from magnetic resonance (MR) imaging. One of the main strengths of the proposed model lies in its use of domain-specific pre-trained weights from knee MRI data that notably outperformed generic or non-knee-specific pretraining approaches in feature representation. By tailoring the pretraining approach to the distinct nature of knee pathologies, MedNet-FS achieved substantially better classification accuracy on both anterior cruciate ligament (ACL) and meniscus tear classification tasks. This performance gap was more prominent in comparison with models initialized using general medical or video-derived features, as shown in Fig. 3A–D. In addition, integrating the generalized end-to-end (GE2E) loss into the MedNet-FS architecture introduced an extra optimization at the embedding level that boosted intra-class compactness and inter-class separability–features crucial for few-shot learning, where samples per class are scarce. The effectiveness of GE2E was amply demonstrated in comparisons of MedNet-FS against its baseline variant (MedNet) on the ACL and meniscus classification tasks, where MedNet-FS achieved statistically significant gains in sensitivity, specificity, and area under the curve (AUC), as seen in Fig. 3c–f. Notably, under support settings of k = 40, MedNet-FS performed at par with, and in some cases outperformed, supervised models trained on corresponding datasets, thus highlighting the potential of FSL frameworks to reduce dependency on large-scale annotations while still providing clinically viable diagnostic accuracy.

The external validation of MedNet-FS on the KneeMRI dataset additionally highlighted its strength and real-world applicability in the clinical environment. When applied to distinguish ACL tears from intact ACLs across the full spectrum of injury presentations, including partially torn ligaments, the model showed only modest effectiveness, with an AUC of 0.58 at k = 40 (Fig. 4a,b), as presented in Table 3. This finding highlights one of the key challenges of using few-shot learning with real-world clinical datasets: the challenge inherent in consistently classifying ambiguous intermediate states of injury, i.e., partial tears, where structural features can be subtle and overlap with healthy and fully torn states. However, when the classification task was modified to include only well-defined categories–intact versus fully ruptured ACLs, in this case, excluding partial tears–the model performance showed significant improvement. In this scenario, MedNet-FS achieved an AUC of 0.62 at k = 40, with balanced sensitivity and specificity (Fig. 4c,d), thus demonstrating its potential utility as a diagnostic tool for well-defined clinical categories. These findings suggest that while the current iteration of MedNet-FS might require additional refinement for borderline or equivocal cases, it is already well-placed as a screening tool or diagnostic aid system for discrete pathologies, i.e., complete ligament ruptures. Such roles could help radiologists and orthopedic clinicians streamline decision-making workflows and increase diagnostic throughput, especially in high-volume or resource-constrained clinical environments^7,29,30.

Situated within the context of the developing literature on few-shot learning in medical imaging, our results add to an expanding corpus of work demonstrating the utility of FSL for disease classification from small amounts of data. Recent publications have successfully implemented few-shot models in a range of clinical specialties, including oncology, ophthalmology, and pulmonary imaging, highlighting the adaptability of FSL approaches to mitigating data paucity^20,31,32. Despite this, relatively few studies have expressly examined few-shot classification with 3D medical imaging data like MRI or CT, where volumetric complexity as well as anatomical variability present distinct challenges^15,33. In this regard, the external data performance of MedNet-FS at an AUC of 0.62 for discriminating intact from completely ruptured ACLs is competitive with or better than the performance described in a number of recent few-shot studies involving analogous imaging modalities. This supports the framework’s promise for more widespread use and shows its viability as an extensible solution for clinical applications featuring high-dimensional imaging inputs and small numbers of annotated examples.

From a systems design standpoint, MedNet-FS presents several opportunities for real-world integration and clinical usefulness. Aside from being a diagnostic tool for musculoskeletal radiologists and orthopedic surgeons, the model can also function as a means of facilitating research in rare or underrepresented knee pathologies for which data collection is necessarily limited. Specifically, its few-shot learning ability can minimize the overhead of manual annotation and enable quick deployment in areas where expert annotation is difficult to obtain. Additionally, the framework can assist research workflows in academic and hospital institutions by being used as a feature extractor or base model for more elaborate multi-stage pipelines. The ability to achieve high performance with limited training data makes MedNet-FS especially suitable for deployment in resource-constrained hospitals or clinics without access to large, labeled datasets or specialized radiological equipment^34,35.

Limitations

Despite the significant advances reflected in MedNet-FS, several limitations need to be addressed before its possible deployment in extensive clinical use. Paramount among these limitations is its reduced performance when dealing with equivocal or intermediate cases, including partial ACL tears. These situations are especially challenging because of the subtle nature of anatomical differences and the overlapping features they share with intact and ruptured ligaments. It is also critical to provide assurance of robust generalizability to heterogeneous populations and multiple scanner types since medical imaging datasets often demonstrate significant variability in acquisition protocols and patient demographics. In addition, the flawless integration into existing clinical information systems and picture archiving and communication systems (PACS) is an important priority to ensure practical feasibility in real-world environments. Without such integration, even artificial intelligence models with high accuracy may face barriers to their deployment in everyday clinical practice.

Future work

Future avenues should involve the enlargement of training datasets for enhancing model diversity, along with the creation of multi-modal data fusion methods that integrate imaging data with clinical notes, radiology reports, or laboratory results. For improving classification success in challenging instances like partial tears, increasing model specialization may be required, which could be accomplished by incorporating segmentation modules or leveraging multi-modal data highlighting localized areas of pathology^36,37. Furthermore, semi-supervised learning methods may be utilized to harness large-scale unlabeled data and further enhance model generalization. The use of explainable AI (XAI) methodologies will also be critical for enhancing model interpretability and clinician trust in automated predictions. Developing intuitive XAI interfaces that emphasize decision-relevant features will foster transparency and facilitate effective human–AI collaboration in the clinical environment. Finally, prospective clinical trials will be required to evaluate the effect of MedNet-FS on diagnostic accuracy, workflow efficiency, and patient outcomes. Through sustained collaboration between AI researchers, radiologists, and orthopedic clinicians, models such as MedNet-FS can be iteratively refined and optimized to address the multifaceted needs of real-world healthcare and unlock the full potential of few-shot learning in contemporary medicine. Additionally, exploring synthetic data augmentation techniques at the embedding level (e.g., SMOTE for feature space oversampling) may further enhance class separation in ambiguous few-shot scenarios.

Conclusion

This study successfully developed and validated MedNet-FS, a 3D few-shot learning (FSL) framework tailored for the classification of knee joint injuries using magnetic resonance (MR) images, effectively addressing the persistent challenge of limited annotated data in medical artificial intelligence. Central to its success was the demonstration that domain-specific pre-training–particularly using knee MRI images–and the integration of generalized end-to-end (GE2E) loss were critical for generating highly discriminative and task-relevant feature embeddings, thereby enabling MedNet-FS to achieve performance levels comparable to supervised learning models even in data-constrained environments. The internal validation conducted on the MRNet dataset established the effectiveness of MedNet-FS in handling few-shot classification tasks for both ACL and meniscus tear conditions, and the subsequent external validation on the KneeMRI dataset further highlighted the model’s clinical viability, particularly in distinguishing fully ruptured ACLs from intact ones. Although classification performance declined when ambiguous partially torn ACLs were included (with the area under the ROC curve, AUC, reaching only 0.58 at k = 40), the model attained significantly improved results–achieving an AUC of 0.62–when the evaluation focused on more clearly separable categories, underscoring the sensitivity of few-shot learning frameworks to class ambiguity and data quality. Importantly, MedNet-FS demonstrated that effective diagnostic learning could be achieved without reliance on large-scale, fully labeled datasets, offering a scalable and rapidly deployable solution for healthcare settings where resource constraints and limited expert availability hinder the adoption of conventional deep learning systems. By reducing the annotation burden and optimizing classification through task-aligned, pre-trained embeddings, MedNet-FS represents a substantial advancement in AI-driven diagnostic tools for musculoskeletal assessment, offering a promising path toward more accessible, efficient, and adaptive medical imaging technologies with the potential to significantly improve patient care.

Author contributions

Xu Lu led the model development, experiments, and manuscript writing. Hongming Lin assisted with implementation and data analysis. Shanhua Sun contributed to data acquisition and medical interpretation. Hua Li supervised the project and provided guidance on the learning framework. Meng Ji reviewed the manuscript and supported project coordination.

Data availability

The datasets used in this study are publicly available from their original sources. The MRNet dataset is available through the source cited in Reference⁹, and the KneeMRI dataset is available through the source cited in Reference¹³. The datasets are not redistributed with this article. The implementation code developed for this study, including data preparation, preprocessing, MedNet-FS model construction, GE2E loss implementation, few-shot evaluation, and performance metric calculation scripts, has been publicly archived on Zenodo and is available at https://doi.org/10.5281/zenodo.20278929.

Declarations

Competing interests

The authors declare no competing interests.