Selection of CNN, Haralick and Fractal Features Based on Evolutionary Algorithms for Classification of Histological Images

The analysis of histological image features for automatic detection of pathologies plays an important role in medicine. Considering that, we proposed a method based on the association of features extracted by multi-scale and multidimensional fractal techniques, Haralick descriptors, and CNN for pattern recognition of colorectal cancer, breast cancer, and non-Hodgkin lymphomas. For feature selection, we applied the ReliefF algorithm to rank the best 50 features and then applied the evolutionary algorithms GWO, PSO, and GA. The classification was made with SVM, K*, and Random Forest algorithms. This strategy allows classifying plenty of feature vectors selected by different algorithms, and consequently, improves the accuracy of the interpretations about the class distinction of histological images. The best combination found was composed of GA and K* algorithms, resulting in 91.06%, 90.52% e 82.01% accuracy for colorectal cancer, breast cancer, and non-Hodgkin lymphomas respectively. The performance obtained by the method indicates that the feature association extracted by different approaches and their subsequent selection and classification presents a potential field for further studies with a high degree of contribution to science.