Endometrial cancer must be caught early to be treated effectively. But identifying it requires distinguishing between three clinically distinct conditions: non-atypical endometrial hyperplasia (NAEH) - a benign thickening that rarely progresses to cancer; atypical endometrial hyperplasia (AEH) - a pre-cancerous lesion where 20-50% of cases can progress to cancer if untreated; and endometrial cancer (EC) itself. These three conditions require very different treatments, so accurate classification is critical.
Transvaginal ultrasound (TVU) is the first-line non-invasive tool for detecting endometrial abnormalities. It is safe, low-cost, and well-tolerated compared to biopsies. However, its accuracy heavily depends on the experience of the sonographer performing the scan. Previous research found accuracy ranging from 82% to 92% between less and more experienced observers examining the same ultrasound images - a troubling degree of variability for a diagnostic tool.
Standard deep learning AI models could potentially standardize TVU interpretation, but they require hundreds or thousands of labeled training images per condition - which are difficult to collect for rare or resource-limited settings. Few-shot learning (FSL) is a machine learning approach that addresses this limitation: it can classify new examples accurately after being shown only a handful of labeled examples per class.
The study's few-shot learning approach used a 3-way 3-shot framework - meaning it was given exactly 3 labeled examples of each of the 3 conditions (9 images total as a "support set") and asked to classify 30 new unlabeled images per condition (90 images total as a "query set"). This mirrors a realistic clinical scenario where labeled expert-confirmed cases are scarce at a new institution.
To compensate for so few training examples, the researchers used dual pretraining with ResNet50 V2 (a powerful CNN architecture). The model was first pretrained on ImageNet - a massive general image database - then fine-tuned further on TVU images of normal uterus and uterine fibroids. This two-stage transfer learning gave the model rich visual feature detection capabilities before it ever saw any NAEH/AEH/EC examples.
The pretrained ResNet50 V2 was used purely as a feature extractor - converting each TVU image into a compact numerical representation (a 64-dimensional vector summarizing the image's visual characteristics). Classification was then performed using a simple k-nearest neighbor (KNN) algorithm: each unknown image was assigned the same label as the majority of its 9 nearest support-set neighbors in feature space, measured by Euclidean distance.
The few-shot learning model (FSL) achieved the best overall performance among all methods tested: overall accuracy of 0.878 and macro precision of 0.882. It outperformed both AutoML models (best: 0.758 accuracy with an Ensemble model using 300 labeled examples per class) and the traditional deep learning model (0.737 accuracy despite being trained on 300 examples per class). This means FSL matched or exceeded conventional approaches using fewer than 1% of the training data.
Crucially, the FSL model was also compared against human sonographers given the same 9 reference images. A junior sonographer (under 5 years' experience) correctly classified only 4 of 30 EC cases and missed all 30 AEH cases. A senior sonographer (over 15 years' experience) correctly identified only 8 EC cases and 10 AEH cases. The FSL model, operating from the same 9 reference images, correctly classified 27 EC and 24 AEH cases - dramatically outperforming both.
For endometrial cancer specifically, the FSL model achieved precision of 0.964, recall of 0.900, and F1-score of 0.931 - the highest values of any method tested. This means it correctly identified 90% of all EC cases while keeping false positives very low. In clinical terms, this translates to fewer missed cancers and fewer unnecessary biopsies on benign lesions.
Transfer learning is the key that makes few-shot learning viable. When ResNet50 V2 is pretrained on millions of general images, it learns to detect universal visual features: edges, textures, shapes, and patterns. When then fine-tuned on uterine ultrasound images, it adapts these general visual skills specifically to the domain of uterine tissue echogenicity, endometrial thickness, and myometrial patterns. These rich learned representations allow meaningful comparison between images even with very few labeled examples.
The KNN classification step works by asking: which labeled images in the support set does this new image most resemble, in the feature space learned by ResNet50 V2? The model computes the Euclidean distance between the feature vector of each query image and all 9 support images, then assigns the label of whichever category is most similar. This simple, intuitive approach is more robust than complex classifiers when only a handful of training examples are available.
Both the DL and FSL models also used Grad-CAM heatmaps for visual explanation - highlighting which regions of each ultrasound image most influenced the classification. This allows sonographers to verify that the AI is focusing on the endometrial lesion region itself rather than irrelevant image artifacts - an important quality check for clinical trust.
The study's most important practical finding is that an AI system can outperform experienced clinical specialists at TVU-based endometrial classification when given the same tiny reference set they are given. This has direct implications for smaller hospitals and clinics where senior sonographers are scarce or unavailable.
Few-shot learning's data efficiency makes it particularly valuable for rare conditions and settings where large annotated datasets are impractical to collect. Rather than requiring each institution to gather hundreds of expert-labeled cases before deploying AI, an FSL model could be set up with as few as 3 reference cases per condition - potentially sourced from a remote expert center rather than requiring local annotation.
Limitations include the modest sample size (33 patients per class in the evaluation set), data from only two hospitals in one geographic region, and the requirement for manual annotation of the region of interest in each TVU image before classification. Future work should validate the approach in larger, geographically diverse cohorts and explore automated lesion detection to reduce the manual preparation step.