hlhong@swu.ac.kr

Our method can be used to the multi-class classification of renal masses and the early diagnosis of renal cancer.

Classification of renal masses as benign angiomyolipoma without visible fat (AMLwvf) and malignant clear cell renal cell carcinoma (ccRCC) in CT images is an important task for early diagnosis of renal cancer. Two masses are known to be similar in intensity in CT images, but it has been also known that ccRCC is more heterogeneous in quantitative texture and AMLwvf is less circular and is more invasive in shape. Thus, we propose a quantitative texture and shape feature classification method for distinguishing AMLwvf from ccRCC in contrast-enhanced (CE) CT images.

Our method was evaluated on a dataset consisting of 80 abdominal CE CT scans from 39 patients with AMLwvf and 41 patients with ccRCC. Scans were acquired on multi-detector scanners at 100s to 120s delays after contrast injection. For each scan, a renal mass was manually marked by a radiologist. From these mass regions, 102 texture features consisting of 7 histogram features, 14 GLCM, 22 GLRLM, and 59 LBP, and 7 shape features consisting of area-perimeter ratio, convex area, eccentricity, major and minor axes lengths, perimeter, and solidity, were extracted. A support vector machine (SVM) and random forest (RF) classifiers were then trained to classify the unseen masses. Experiments were cross validated by leave-one-out method and the training images were 250-fold augmented by random rotation, translation, and scaling to avoid the overfitting. The classification performance of the texture feature alone was 73.8% for both SVM and RF, while the performance of the texture and shape features was improved to 78.8% and 76.3% for SVM and RF, respectively.

Our texture features provide the distinctive patterns of renal masses which are difficult to visually distinguish. Our shape features further improve the classification performance by reflecting the shape difference between two masses to the classifier. Throughout our experiments, it is quantitatively confirmed that there is a meaningful difference in texture and shape between two renal masses.