Preliminary study on stromal nodule detection using b-mode ultrasound images and deep learning

Benign prostatic hyperplasia (BPH) is a common condition among aging men and requires accurate diagnostic tools. Transrectal ultrasonography (TRUS) is widely used for BPH detection but often lacks accuracy. In this study, we investigated a deep learning approach for stromal nodule detection in ex-vivo prostate specimens exhibiting BPH using B-mode ultrasound imaging. B-mode data were acquired from human prostate specimens, with approximately 200 ultrasound frames per sample. Pathologist-provided annotations distinguishing stromal and non-stromal regions were aligned with corresponding ultrasound images and used as ground truth. The Segment Anything Model (SAM) was trained in a supervised manner to segment stromal versus non-stromal tissue, using an 80/20 patient-level train/test split for evaluation. Preliminary results demonstrate the feasibility of transformer-based segmentation for stromal nodule detection in B-mode ultrasound, achieving a Dice score of 0.76. This approach may enhance the diagnostic capability of ultrasound imaging for BPH detection, offering a non-invasive and accurate method that could improve patient outcomes and enable more personalized treatment planning.