A Deep Learning Approach to Tumour Identification in Fresh Frozen Tissues

The demand for pathology services are significantly increasing whilst the numbers of pathologists are significantly decreasing. In order to overcome these challenges, a growing interest in faster and efficient diagnostic methods such as computer-aided diagnosis (CAD) have been observed. An increase in the use of CAD systems in clinical settings has subsequently led to a growing interest in machine learning. In this paper, we show the use of machine learning algorithms in the prediction of tumour content in Fresh Frozen (FF) histological samples of head and neck. More specifically, we explore a pre-trained convolutional neural network (CNN), namely the AlexNet, to build two common machine learning classifiers. For the first classifier, the pre-trained AlexNet network is used to extract features from the activation layer and then Support Vector Machine (SVM) based classifier is trained by using these extracted features. In the second case, we replace the last three layers of the pre-trained AlexNet network and then fine tune these layers on the FF histological image samples. The results of our experiments are very promising. We have obtained percentage classification rates in the high 90s, and our results show there is little difference between SVM and transfer learning. Thus, the present study show that an AlexNet driven CNN with SVM and fine-tuned classifiers are a suitable choice for accurate discrimination between tumour and non-tumour histological samples from the head and neck.