Impact of Image Pre-Processing and Optimizers on U-Net Segmentation of Prostate Cancer in MRI

Prostate cancer remains among the most frequently diagnosed cancers in men and is a major issue in achieving timely and accurate diagnosis. Magnetic Resonance Imaging (MRI) is widely applied because it is capable of picking up high-resolution anatomical information, but cancer regions must be manually segmented, which is time-consuming and prone to the variability of experts. This study proposes an automated segmentation algorithm using the U-Net deep learning model with image pre-processing techniques to address these deficiencies. Median filtering was done to remove salt-and-pepper noise, followed by brightness enhancement to improve the intensity contrast of images. The pre-processed images were used to train a U-Net model for prostate cancer segmentation. The Dice Similarity Coefficient (DSC) metric was used to evaluate the segmentation accuracy. Three optimizers, Adam, RMSprop, and Adagrad, were tested. All of them were trained between 10 and 100 epochs. Adam optimizer obtained the highest segmentation performance at epoch 90 with the highest DSC value of 0.9907, while RMSprop and Adagrad produced 0.9888 and 0.9655, respectively. Pre-processing raised the mean DSC from 0.8206 to 0.8733, confirming its impact on image quality enhancement. Overall, the proposed method demonstrates high accuracy and reliability, offering a practical solution to support radiologists in prostate cancer diagnosis and treatment planning.