An Efficient Deep Learning Model for Prostate Cancer Diagnosis

Prostate cancer is one of the most widespread and mortal cancers among men worldwide. Pathological evaluation is essential for the definitive diagnosis of prostate cancer, the symptoms of which are not evident during the early disease. During the pathological evaluation to determine the severity of cancer, Gleason grading, which is difficult to assign, is time-consuming, and the result may vary depending on the pathologist's experience. In order to make a fast and objective evaluation, deep learning systems have been recently used to automate prostate cancer diagnosis and grading. The inherent difficulties of detecting Gleason grades can be overcome with deep learning architectures, which present fast and precise classification rates. The performance of the methods proposed for Gleason grading in the literature is low, especially in multi-class problems, due to their model structures. Some mechanisms for better learning of critical representations are needed to determine the Gleason grading accurately. Therefore, in this study, a novel deep learning model called Eff4-Attn is proposed for detecting and grading prostate cancer. In this model, combining the Efficient Channel Attention (ECA) mechanism with EfficientNet-B4 distinguishes the tissues in the prostate cancer pathology images much better. In addition, the effect of data distribution in classes in the public DiagSet dataset on performance was investigated. Experimental studies have been carried out to perform three classification tasks for various magnification ratios: 8-class classification, 4-class classification, and 2-class classification for diagnosing prostate cancer and determining cancer severity, and promising results have been obtained. At 40x magnification ratio, the best result was achieved with 96.18% test accuracy in the 2-class classification task, that is, in detecting the presence of cancer. In determining the severity of cancer, at a 40x magnification ratio, a test accuracy of 94.86% was obtained. The proposed deep and light model has considerable potential to assist pathologists in detecting the characteristics of Gleason grades. © 2013 IEEE.