Cascaded Architecture for Segmenting Prostate Cancer Lesions in Biparametric MRI

Prostate cancer (PCa) is one of the most common cancers in men. Early detection and appropriate treatment are crucial for managing the disease. Clinical approaches can assess prostate cancer status based on the PSA level in the blood. However, it is difficult t o determine t he exact location of the cancerous lesions in the prostate. Physicians typically use biparametric magnetic resonance imaging (bpMRI) to identify suspicious areas, but its complexity and variability make interpretation challenging. This study investigated a convolutional neural network (CNN) approach for for segmentation of prostate cancer lesions in bpMRI. We used a cascaded two-stage approach with two SegResNet models. The first model helps t o narrow down the area for segmenting cancerous lesions. The second model segments cancerous lesions within the prostate area. The model processes 3D bpMRI images with T2W, DWI and ADC sequences. The model was trained on the PI-CAl 2022 dataset. The performance of the model was evaluated externally using the Prostate-l58 and our PSU-mpMRI datasets. Our model showed higher performance compared to a single SegResNet baseline. We achieved a Dice score of 0.5334, 0.4221 and 0.3588 for the PI-CAl 2022, Prostate-l58 and PSU-mpMRI datasets, respectively. Our model demonstrated generalisability towards external data.