Clinically Significant Prostate Cancer (csPCa) segmentation/classification using deep learning on multi-parametric Magnetic Resonance (MR) Images

Abstract

Prostate cancer (PCa) is one of the most common types of cancer found in males that affects the prostate gland. The recent advancement has led to the use of MRI images that provide an accurate alternative for PCa characterization and detection. The use of MRI often suffers from significant inter/intra observer variability among radiologists. However, recent advancements in Computer Aided Detection (CAD) and computer vision for prostate cancer lesion segmentation have shown great improvement. Despite these advancements, characterizing prostate cancer remains a significant challenge. The timely detection and classification of prostate cancer is crucial for early diagnosis, thus forming the problem statement for this study. This study presents a single-shot approach to perform multi-class segmentation on MRI images to provide with PCa lesions. PI-CAI public and development dataset was used for segmentation and classification, the research involved the use of variants of the popular U-Net model, the 3D U-Net and MutliResU-Net for multi-class segmentation that showed remarkable progress for multi-class segmentation. 3D U-Net particularly produced more accurate segmentations and classification for higher Gleason scores as well with an overall sensitivity of 83% on the test set, the MultiRes U-Net also showed good segmentations only with a Sensitivity of 64% showcasing the potential of these approaches for multi-class segmentation for prostate cancer. The implications of this work are significant and lay the groundwork for effective multi-class segmentation contributing to impactful future innovations that can adapt AI for routine clinical settings.