AUA 2023: Virtual Renal Mass Biopsy: Predicting Renal Tumor Histology on Abdominal CT Images using Machine Learning

( On the first day of the annual American Urological Association meeting in Chicago, Dr. Abhinav Khanna explored the association between advancing technology and renal tumors. Specifically, Khanna et. al from Mayo Clinic looked to use a machine learning mechanism to diagnose benign vs malignant renal tumors. Treatment for any renal mass results in partial or radical nephrectomies depending on the severity of disease. Benign renal tumors can only be diagnosed through biopsies which are invasive and lead to higher false negative rates. Given that 20% of renal masses are non-malignant, these operations may be unnecessary. Dr. Khanna stresses in his presentation that 1 out of 3 partial nephrectomies are unnecessary. Therefore, finding alternative and more efficient ways to distinguish malignant vs nonmalignant is of great importance for patients who may be able to avoid invasive surgery if ultimately successful.

Dr. Khanna et. al. gathered CT scans from 609 patients which included both nonmalignant and malignant renal tumors. Of those 609 patients only 168 CT images were used in the training set to train this algorithm. The team then applied an artificial intelligence algorithm that has been previously used for kidneys, cysts, and tumor area. The algorithm used all different segments of the CT scan to fully make a 3D model of the kidney and tumor (Figure 1). They also assessed the top radiomics (medical images and using data characteristics algorithms) and found they are at most 60% accurate, a subpar amount.


Figure 1. Shows the CT scans and how they segmented kidneys and kidney tumors to train the machine learning mechanisms.

The team at Mayo clinic found that the algorithm is 0.90 accurate in detecting nonmalignant and malignant tumors. In addition, the algorithm also had a .84 sensitivity and 0.97 specificity (Figure 2). Results here indicate that the machine learning algorithm distinguishes oncocytoma vs malignant renal neoplasm with promising accuracy. At the end of his presentation, Dr. Khanna presented three excellent future directions. The first direction is to integrate the algorithm into the radiology workflow to provide better decision support for clinicians. The second direction is to conduct external validation tests to determine if the algorithm is effective in different settings. Lastly, Dr. Khanna suggested testing the algorithm on indolent versus aggressive malignancies to determine its effectiveness in treating different types of cancers.


Figure 2. shows a confusion matrix of the 168 reviewed images and its correct or incorrect classification of masses.

The presentation from Dr. Khanna sparked many conversations in the crowd. One audience member asked how the institution was able to standardize the CT images since the type of CT scans done can vary in quality and what it may show. Dr. Khanna replied that they standardized contrast enhancements which data scientists were able to account for any variability. An audience member then asked how reliable the results are given that pathologist struggle to diagnose biopsy samples. Dr. Khanna replied, “I wholeheartedly agree that’s why we didn’t use biopsy specimens to train the model, we used tumors that were cut out”, which clarifies methods used in training the program.

Dr. Khanna et. al provides a potentially great solution to correctly identifying oncocytoma and malignant renal masses using artificial intelligence. A great example of how advanced technology can be utilized in medicine. If proven to achieve external validity and can correctly identify indolent vs malignant tumors, this machine learning technology can better guide physicians to treat their patients.

Presented by: Abhinav Khanna MD, Mayo Clinic, Rochester, MN

Written by: Paul Piedras, B.S., University of California, Irvine, @piedras_paul on Twitter, during the 2023 American Urological Association (AUA) Annual Meeting, Chicago, IL, April 27 – May 1, 2023