Hybrid models for classifying histological images: An association of deep features by transfer learning with ensemble classifier

Abstract

The use of a convolutional neural network with transfer learning is a strategy that defines high-level features, commonly explored to study patterns in medical images. These features can be analyzed via different methods in order to design hybrid models with more useful and accurate solutions for clinical practice. In this paper, a computational scheme is presented to define hybrid models through deep features by transfer learning, selection by ranking and a robust ensemble classifier with five algorithms. The obtained models were applied to classify histological images from breast, colorectal and liver tissue. The strategy developed here allows knowing important results and conditions to improve models of computer-aided diagnosis, even exploring classic CNN models. The features were defined using layers from the AlexNet and ResNet-50 architectures. The attributes were organized into subsets of the most relevant features and submitted to a k-fold cross-validation process. The best hybrid models were obtained with deep features from the ResNet-50 network, using distinct layers (activation_48_relu and avg_pool) and a maximum of 35 descriptors. These hybrid models provided 98.00% and 99.32% of accuracy values, with emphasis on histological images of breast cancer, indicating the best solution among those available in the specialized Literature. Also, these models provided more relevant results for classifying UCSB and LG datasets than regularized techniques and CNN architectures, exploring data augmentation or not. The computational scheme with detailed information regarding the main hybrid models is a relevant contribution to the community interested in the study of machine learning techniques for pattern recognition.