Gynosight AI: A Trustworthy Explainable Ultrasound Assistant for Obstetrics and Gynaecology

Ultrasound is the main imaging test in obstetrics and gynaecology, used to monitor fetal growth and to examine the ovaries and uterus. Existing artificial intelligence tools each handle only one condition, give an answer even when the image is blurred or unfamiliar, and show heatmaps that are judged only by eye. This project proposes GynoSight AI, a trustworthy explainable ultrasound assistant with four modules: fetal growth (head circumference measurement, with an amniotic fluid check from the amniotic fluid index entered by the doctor), ovarian mass classification, polycystic ovary syndrome (PCOS) detection and uterine fibroid detection. The modules use DenseNet121, EfficientNet-B0 and U-Net with transfer learning, and Grad-CAM++ highlights the region behind each finding. A trust gate combines image quality, Monte Carlo Dropout uncertainty, an energy-based familiarity score and a measurable heatmap plausibility score. Conformal risk control sets the acceptance threshold so that the error rate among accepted cases stays below a chosen limit, and every rejected case is referred to a gynaecologist with a plain-language reason. Phase–I covers the literature survey, gap analysis, requirements, system design and evaluation plan; implementation and testing on public datasets are planned for Phase–II.