Semi-Automatic 3D Medical Image Segmentation Using UNet with Five-Point Extreme Guidance

Breast cancer remains the leading cause of cancerrelated mortality among women worldwide, with millions of new cases reported annually. Advances in early detection and treatment, particularly through mammography and tomographic imaging, have contributed to improved survival rates. This work presents a work-in-progress semi-automatic segmentation framework for breast lesion annotation in Digital Breast Tomosynthesis (DBT), leveraging a 3D UNet with spatial priors based on five-point user input. The method combines expert-in-the-loop interaction with deep learning, encoding four extreme boundary points and one central point as a Gaussian heatmap to guide volumetric segmentation. Evaluated preliminarily on a cohort of 81 patients from the multicenter P.I.N.K. study, the model shows promising qualitative localisation while significantly reducing annotation effort. Although quantitative metrics are not yet reported, the system is designed for clinical adaptability, with planned extensions including real-time user interaction and benchmarking against interactive segmentation baselines.