Prostate Cancer Classification using Convolutional Neural Networks

Abstract: In 2012 prostate cancer was the second most common cancer diagnose for men. The diagnosis is confirmed by pathologists doing ocular inspection of prostate biopsies and the specimens are classified according to the Gleason grading system. The main goal of this thesis is to automate the classification using Convolutional Neural Networks (CNN). With the introduction of Convolutional Neural Networks the field of pattern recognition broadened. The classical way of designing and extracting hand-made features for classification is substantially different to letting the computer itself decide which features are of importance, the new approach was enabled by CNNs. This together with groundbreaking results on benchmark image sets has made CNNs a well-used method in pattern recognition. In this thesis a CNN with small convolutional filters has been trained from scratch using stochastic gradient descent with momentum. The error rate for the CNN is 7.3%, which is significantly better than previous works using the same data set. Since good results were obtained even though the data set were rather small, the conclusion is that CNNs are a promising method for this problem.