AI-Based Multi-Organ Segmentation in Gynecologic Laparoscopy: Comparative Evaluation of Deep Learning Architectures...

Recent advances in deep learning have shown considerable promise in automating anatomical segmentation for laparoscopic gynecological surgery. However, segmenting small and morphologically variable structures such as the ovary and fallopian tube remains a significant challenge. In this study, we conducted a comparative evaluation of three convolutional neural network models-U-Net, Segformer, and MANet-on uterus, ovary, and fallopian tube segmentation using 381 laparoscopic images from 21 patients undergoing gynecologic procedures. U-Net demonstrated the most consistent performance, achieving Dice Similarity Coefficients (DSC) of 0.9068 ± 0.0636 for the uterus, 0.2699 ± 0.1273 for ovaries, and 0.2660 ± 0.1187 for fallopian tubes. MANet exhibited the highest precision for uterus segmentation (0.9182 ± 0.0594), while Segformer provided balanced specificity across structures. Despite these strengths, all models exhibited substantial performance degradation when segmenting smaller organs, with a 70.6% reduction in DSC between uterus and fallopian tube segmentation. These results highlight persistent technical challenges in laparoscopic image analysis, particularly for small or ambiguous structures. These findings provide an initial benchmark highlighting both the promise and the limitations of current deep learning approaches for laparoscopic gynecologic image analysis and emphasize the need for further research to improve segmentation of mobile, low-contrast structures under real-world surgical conditions.