Histopathology image classification based on ensembles of highly compact features

In this paper, we present a method for extracting highly informative yet compact features from histopathological images, which at the same time enable their classification with high accuracy. These types of concise features are very valuable due to the ever-increasing size of datasets as well as the sheer size of histopathology scans. To meet these requirements, we propose new methods for extracting features from the combinations of inner layers of deep neural networks DenseNet-161 and ResNet-152, as well as DINO derived from vision transformers. This is followed by the construction of the subspaces with the help of the principal, as well as neighborhood component analysis. This method creates new features: PCA/NCA-DenseFeats, PCa/NCA-ResFeats, and PCA/NCA-DINO. We then further improve their performance through domain-trained extraction models. The collected results show that the classification accuracy of these features becomes stable even after several-fold reduction of the dimensionality of the baseline features. This means that there is no need to use long representation vectors, and features even several times shorter can be used for classification with almost no information loss. In addition, we propose the multi-architecture ensemble of classifiers orchestrated by the error-correcting output codes method, which utilizes features extracted from three different domain-finetuned models with different architectures: ViT Small, ResNet-152, and DenseNet-161. The proposed solution allows for an increase in accuracy by up to almost 7pp compared to the baseline. The classification vector consumes only 180 float16 values, which is significantly smaller than many other published methods. This reduces the memory required to store features and enables faster classification of these representations. The proposed method was evaluated on the BreakHis breast cancer dataset (achieving 95.39% at 40X, 92.07% at 100X) and the Kather colorectal cancer dataset (97.1%), demonstrating its generalizability and applicability.