Augmenting endometriosis analysis from ultrasound data with deep learning

This retrospective pilot study evaluated the feasibility of classifying endometriosis from transabdominal and transvaginal ultrasound images using five pretrained convolutional neural networks (Xception, Inception-V4, ResNet50, DenseNet, EfficientNetB2) with binary cross-entropy loss, ImageNet initialization, and 5-fold cross-validation where all images from a given patient stayed within the same fold. Using 100 subjects (50 ultrasound-normal and 50 laparoscopically/laparotomically diagnosed endometriosis) and standardized preprocessing with image cropping, resizing, and augmentation, the DenseNet model achieved an AUC of 0.90 and accuracy of 80% on “Test 1” (augmented-image cross-validation) and improved performance on “Test 2” video-derived data (AUC 0.90, accuracy 84%) where augmentation was not used. The authors explicitly frame this as a pilot study and rely on retrospectively collected data from a single hospital with performance reported at the image level, not per-lesion or per-patient outcomes. This paper is centrally about endometriosis — deep learning classification of endometriosis from ultrasound data (including transabdominal and transvaginal images and video).

Read from the paper's body, not the abstract. Not a substitute for reading the paper. No clinical advice. How this works