Textural and deep learning methods in recognition of renal cancer types based on CT images

Abstract

Recent advancements in deep learning have opened new prospects in many areas of research. Especially interesting field is biomedical image analysis, where plenty of problems wait for efficient solution. The aim of this work is to develop new approaches to recognition of different types of renal cancer on the basis of Computed Tomography (CT) imaging. Two different directions will be investigated. One uses the texture descriptors of the images to define the diagnostic features. They are next combined with support vector machine responsible for final recognition and classification. The second applies deep learning approach using different configurations of Convolutional Neural Networks. The experimental research for both textural and deep learning approaches was conducted on real world dataset of CT scans consisting of eight types of renal cell carcinomas. The proposed structures of predictive system were able to achieve the level of accuracy around 90% for complex and unbalanced datasets.