Accurate classification of endometrial lesions is essential for appropriate patient management. The spectrum ranges from entirely benign conditions - such as endometrial polyps (EP) and submucosal myomas (SM, benign fibroids) - through precancerous endometrial hyperplasia (EH) and the higher-risk atypical hyperplasia (AH), to frank endometrial carcinoma (EC). Each category warrants dramatically different clinical responses, from watchful waiting to immediate surgical intervention.
Hysteroscopy allows direct visual examination of the uterine cavity and is widely used to evaluate these lesions. However, accurate classification during hysteroscopy relies heavily on the gynecologist's experience and visual acuity. Studies have documented that even experienced gynecologists show considerable variation in their diagnostic accuracy, particularly in distinguishing between precancerous and malignant categories, which can look visually similar.
This study asked whether a deep learning model - specifically the well-established VGGNet-16 architecture - could learn to classify hysteroscopy images across these five categories with accuracy comparable to or better than trained gynecologists, and whether the model could serve as a useful diagnostic aid when used alongside clinician judgment.
The dataset consisted of 1,851 hysteroscopy images from 454 patients across the five categories: endometrial hyperplasia (EH), atypical hyperplasia (AH), endometrial carcinoma (EC), endometrial polyp (EP), and submucosal myoma (SM). This represents a substantial clinical dataset, though the number of images per category varied, with some rare conditions having fewer examples than others.
To address class imbalance and increase the effective size of the training set, data augmentation techniques were applied - artificially expanding the dataset by creating variations of existing images through transformations such as rotation, flipping, brightness adjustment, and zooming. After augmentation, the training set expanded to 6,478 images, more than tripling the original size. This is standard practice in medical image deep learning, where collecting large labeled datasets is expensive and time-consuming.
VGGNet-16 is a convolutional neural network architecture developed at Oxford that won the ImageNet large-scale visual recognition challenge in 2014. The "16" refers to its 16 weight layers. The model was initialized with weights pre-trained on ImageNet (a massive dataset of natural photographs) and then fine-tuned on the hysteroscopy images - a strategy called transfer learning that allows models to leverage general visual feature recognition for specialized medical applications with relatively smaller datasets.
In the 5-category classification task, the VGGNet-16 model achieved an overall accuracy of 80.8%. In the 2-category task (benign versus premalignant or malignant), which is the most clinically critical distinction, the model achieved 90.8% accuracy. These performance levels establish the model as a strong baseline for automated endometrial lesion classification.
Three gynecologists evaluated the same test images independently. Their individual 5-category accuracies were 72.8%, 69.2%, and 64.4% - all substantially below the model's 80.8%. This is a significant finding: the deep learning model, trained purely on image data without any clinical context, outperformed practicing clinicians on the same standardized image set. The model's advantage was most pronounced in distinguishing between the more visually similar lesion categories.
The difference between the model's accuracy (80.8%) and the best-performing gynecologist (72.8%) is approximately 8 percentage points - a meaningful clinical margin. Across thousands of hysteroscopy procedures performed annually, an 8-percentage-point improvement in diagnostic accuracy could translate to substantially fewer missed early cancers and unnecessary interventions for benign disease.
A key secondary experiment tested whether gynecologists performed better when they could see the model's predictions alongside the hysteroscopy images. Before seeing the AI output, the three gynecologists' accuracies ranged from 64.0% to 69.2%. After receiving AI assistance, their accuracies improved to a range of 72.8% to 78.4% - gains of approximately 8 to 11 percentage points per clinician.
This human-AI collaboration result is arguably the most clinically important finding in the study. It demonstrates that the model does not replace the clinician but rather enhances their performance. Clinicians likely benefited from the model's output as a second opinion that prompted them to reconsider initial judgments and notice visual features they might otherwise have overlooked.
The improvement was consistent across all three gynecologists regardless of their baseline performance level, suggesting that AI assistance provides broad benefit rather than only helping less-experienced practitioners. Notably, when aided by the model, the weakest-performing gynecologist reached accuracy levels close to the pre-assistance accuracy of the strongest performer, effectively narrowing the performance gap between clinicians of different experience levels.
Transfer learning is central to why deep learning works in medical imaging despite relatively small datasets. VGGNet-16 pre-trained on ImageNet has learned to detect fundamental visual features - edges, textures, shapes, color gradients - from millions of natural images. These low-level features are surprisingly transferable to medical images, and only the higher-level, task-specific feature recognition needs to be learned from the hysteroscopy images.
Without transfer learning, training a 16-layer deep network from scratch on 1,851 images would likely result in overfitting - a situation where the model memorizes the training data rather than learning generalizable patterns. Pre-training prevents this by initializing the network in a well-regularized state that already understands basic visual composition. Fine-tuning then adapts these representations to the specific visual characteristics of endometrial lesion categories.
The data augmentation strategy used in this study illustrates another important technique for working with limited medical datasets. By generating geometrically and photometrically transformed variants of existing images, augmentation teaches the network that the same lesion type can appear in different orientations, under varying lighting conditions, and at slightly different scales - critical robustness properties for a system deployed across different hysteroscopes and operators.
This study makes a compelling case for AI-assisted hysteroscopic diagnosis of endometrial lesions. The model's superior stand-alone accuracy compared to experienced gynecologists, combined with its ability to boost human performance when used as decision support, positions it as a practical tool that could be integrated into hysteroscopy workflows rather than replacing clinical judgment entirely.
The 2-category accuracy of 90.8% for the clinically critical benign versus premalignant or malignant distinction is particularly promising. At this accuracy level, the model could serve as an effective screening tool to flag cases warranting biopsy or closer examination, while allowing operators to confidently proceed with simpler management for clearly benign cases.
Limitations include the retrospective single-center design, the relatively modest dataset size, and the absence of external validation on images from other institutions. Future research should evaluate the model's performance on images from different hysteroscopes and clinical settings, and explore whether performance remains strong across diverse patient populations. Real-time integration into hysteroscopy systems and prospective clinical trials will ultimately determine the system's role in standard of care.