Histological Image Classification using Deep Features and Transfer Learning

A major challenge in the automatic classification of histopathological images is the limited amount of data available. Supervised learning techniques cannot be applied without some adjustment. We compare two common techniques to deal with limited domain data: using deep features and fine-tuning convolutional neural networks (CNN). We examine the following state-of-art CNN models: SqueezeNet-v1.1, MobileNet-v2, ResNet-18, and DenseNet-201. We demonstrate that using feature vectors that are extracted from one of the four CNN models with a classical support vector machine (SVM) for training and testing can lead to higher accuracy on publicly available datasets: Warwick-QU, Epistroma, BreaKHis, multi-class Kather, than previously published results. Similar results can be obtained with fully fine-tuning the aforementioned CNN models. We also study the effectiveness of block-wise fine-tuning of two models (i.e., SqueezeNet-v1.1 and ResNet-18) and show that it is not necessary to fully fine-tune leading to savings in time and space.