Machine learning based quantitative texture analysis of CT images for diagnosis of renal lesions

Abstract

Renal cancer is among the prominent causes of an upsurge in death rates around the globe that can be minimized by premature diagnosis. Computed tomography (CT) is the first-line imaging technique employed for analysis of renal lesions in clinical practice. Moreover, it is difficult for the radiologist to analyze a large number of medical images which may lead to misclassification of renal lesions that eventually results in cancer evolution or needless chemotherapy. Thus, there is a requirement to propose machine learning based techniques for classification of renal tissue based on their texture, and these techniques can be considered as a promising second opinion for the clinicians to minimize false-negative results. The present study aims to estimate the diagnosis performance of machine-learning based quantifiable texture examination of CT images to distinguish renal lesions. Texture analysis is performed to quantify the renal parenchyma tissues by employing different texture models, and the paper suggests a feature vector for the categorization of benign and malignant renal lesions based on their texture analysis. Further, dimensionality reduction is performed to recognize the discriminative features using distinct statistical, similarity and information knowledge-based techniques. The effectiveness of the resultant ranked feature vector is evaluated by using a Support vector machine (SVM), naive Bayes, Sequential minimal optimization (SMO), random forest, and K nearest neighbour (KNN) classifiers. The classification accuracy equivalent to 84% is observed by KNN with 5-fold cross-validation. Additionally, the practical implications, generalizability issues along with certain open challenges are highlighted that may open avenues for further research.