Convolutional Network Identifies Cancerous Regions in Ultrasound

A single-center prospective study (July 2022–February 2023) at Sun Yat-sen University trained CNNs (FCN-101, DeepLabV3) on 386 registered high-frequency ultrasound images and corresponding biopsy whole-slide images from 105 patients to localize breast cancer regions. FCN-101 achieved higher pixel accuracy (86.91%) and Dice (77.47%) than DeepLabV3, demonstrating pixel-level concordance with histopathology and potential for biopsy guidance.