Deep learning could potentially assist healthcare providers with the evaluation of small renal masses detected on certain contrast-enhanced CT exams, according to a new study published in the American Journal of Roentgenology. The AI’s performance excelled with some CT phases but not others.
“Dynamic CT can provide useful diagnostic information for the differentiation of small solid renal masses,” wrote lead author Takashi Tanaka, department of radiology at Okayama University Hospital in Japan, and colleagues. “However, despite careful evaluation of images by experienced radiologists, evaluation can be subjective and affected by each radiologist's experience to a certain extent. In addition, some incidentally detected small renal masses … are unable to be clearly determined to be benign or malignant on the basis of imaging alone; they can ultimately be shown to be benign only after definitive surgical intervention.”
The authors aimed to see if deep learning techniques could help improve the differentiation of small solid renal masses, providing a new sense of clarity to an often challenging process. Their retrospective study included 168 solid renal masses from 159 patients who underwent CT imaging from 2012 to 2016. All masses were 4 cm or smaller, and the CT examinations were all either unenhanced, corticomedullary phase (CMP), nephrogenic phase or excretory phase.
Tanaka et al. divided the dataset into five datasets, with four sets being used for the augmentation and training of a convolutional neural network (CNN). The fifth and final dataset was set aside for testing. Overall, the deep learning model had a much higher area under the ROC curve with the CMP CT images than other phases. The accuracy was also highest, 88%, with the CMP CT images. The accuracy of the excretory phase was 85%, and then it dropped significantly for other phases.
As Tanaka and colleagues noted in their analysis, the deep learning model may have been effective for the CMP CT images, but it was “not as good” for other phases.
“This study showed that deep learning with a CNN can support the evaluation of small solid renal masses with acceptable diagnostic performance,” the authors concluded. “Both the CNN model and the reviewing radiologists concluded that CMP CT images enabled better differentiation of small solid renal masses (≤ 4 cm) than did other phase images on dynamic CT.”