Combining texture analysis and deep learning in renal tumour classification task

Current advances in Computer Vision allowed solving the vast majority of tasks in this field. However, the analysis of medical imaging still requires specialized algorithms, usually employing a lot of expert knowledge. This research aims to create an automated computer system able to differentiate 8 types of renal cancer types. Such a system can contribute to earlier diagnosis and better prioritization of patients waiting for medical consultation. The multi-feature ensemble system integrates 2 distinct groups of feature generation and classification: texture analysis and deep learning. In the first approach, three independent methods of feature generation: GLCM, SFTA, and Unser algorithm have been used. Their results are then passed to the SVM classifier. The second group utilizes transfer learning applied to pretrained deep models, developed within the ILSVR challenge. Combining both individual approaches with weighted voting yields results reaching 93.6 % of weighted F1-score evaluated in 10-fold cross validation mode.