Optimized Light-Weight Convolutional Neural Networks for Histopathologic Cancer Detection

Cancer is one of the leading causes of death across the world, and early detection of cancer will greatly increase the chances of successful treatment. Histopathology diagnosis of cancer is crucial because most diagnoses of malignancy are made by pathological confirmation. In Japan, lack of pathologist has become a serious problem. The lack of pathologists to diagnose the huge amounts of pathological images has resulted the late of medical diagnosis thus delaying the necessary treatment for the patients. Recently, there has been extensive research work on how to incorporate artificial intelligence in the diagnosis process. For example, using convolutional neural networks (CNN) for histopathologic cancer detection can not only speed up diagnosis but also can increase the diagnosis accuracy. In this study, we consider EfficientNet-B6 CNN model for histopathologic cancer detection and apply different activation functions as well as gradient descent optimization algorithms, to analyze its effects on diagnosis accuracy. EfficientNet-B6 is a light-weighted model proposed by Google, an easy-to-deploy-model and works efficiently in classification and detection tasks. Patch-Camelyon benchmark dataset that consists of non-cancerous and cancerous histopathologic images is used in this study. It was found that the top accuracy of this method is 97.94% when it is trained by EfficientNetB6 with Rectified Adam as the optimizer and Mish as the activation function.