Automated prostate multi‐regional segmentation in magnetic resonance using fully convolutional neural networks


Magnetic resonance imaging (MRI) plays a pivotal role in detecting and diagnosing prostate cancer (PCa), addressing some liabilities of traditional methods like prostate-specific antigen (PSA) testing and blind biopsy. However, the diagnostic workload for radiologists remains substantial. Automating the analysis of the prostate gland using advanced computer-assisted detection applications could streamline the identification and scoring of lesions, aiding in prostate and tumor volume calculations. Additionally, differentiating between various prostate zones is crucial for advancing clinical assessment tools, especially in evaluating aggressive cancer. This work proposes a reliable CNN-based automatic prostate multi-regional segmentation method validated across different institutions, countries, and continents. Using a dataset of 243 T2-weighted prostate studies across seven countries, radiologists trained and tested a U-Net-based model (a type of CNN architecture) with manual delineations of prostate regions as reference. The model yielded promising Dice similarity coefficient (DSC) scores for various prostate regions across 120 test studies, showcasing consistency across different equipment and regions. The results held consistency across different manufacturers or continents, exhibiting no statistically significant disparities. This study successfully proves that accurate prostate multi-regional T2-weighted MR images automatic segmentation can be attained through U-Net like CNN, showcasing adaptability across a broad clinical spectrum with varied equipment, acquisition configurations, and population.