ColorNephroNet: Kidney tumor malignancy prediction using medical image colorization

Abstract

Renal tumor malignancy classification is one of the cru- cial tasks in urology, being a primary factor included in the decision of whether to perform kidney removal surgery (nephrectomy) or not. Currently, tumor malignancy prediction is determined by the radiological di- agnosis based on computed tomography (CT) images. However, it is estimated that up to 16% of nephrec- tomies could have been avoided because the tumor that had been diagnosed as malignant, was found to be be- nign in the postoperative histopathological examination. The excess of false-positive diagnoses results in unnec- essarily performed nephrectomies that carry the risk of periprocedural complications. In this paper, we present a machine-aided diagnosis system that predicts the tumor malignancy based on a CT image. The prediction is performed after radiological diagnosis and is used to capture false-positive diagnoses. Our solution is able to achieve a 0.84 F1-score in this task. We also propose a novel approach to knowledge transfer in the medical do- main in terms of colorization based pre-processing that is able to increase the F1-score by up to 1.8pp.