A novel solution of using deep learning for prostate cancer segmentation : enhanced batch normalization

Deep learning has not been successfully implemented in the past with accurate segmentation of prostate on Magnetic Resonance (MR) image in nerve sparing prostate surgery. This was mainly due to the disease-specific change in shape, boundary of gland, and complexity in separating surrounding tissues. This research aims for accurate segmentation of prostate on MR images by combining multi-level features for decreasing the processing time of the process in prostate surgery. The proposed system consists of a deep neural network that extracts high-level and low-level features of MR images, and a propagation technique that combines the extracted features thus increasing the segmentation accuracy and reducing the time required for segmentation process. Accuracy is calculated using dice similarity coefficient and performance is calculated with total execution time of the datasets. The results show improved performance with 2.11 s against 2.29 s. In addition, the overall accuracy is improved to 95.3%% against 92.76% for the MR prostate segmentation. The proposed system focuses on automating the prostate segmentation in MR images with enhanced accuracy, and thus assisting prostate surgeries and disease diagnosis. This study solves the issues related to prostate shape recognition and prostate localization and improves the segmentation accuracy and performance.