AI Helps Radiologists Accurately Distinguish Malignant Vs. Benign Renal Masses

Friday, Dec. 04, 2020

By Nick Klenske



An RSNA researcher has developed an artificial intelligence-based (AI) predictive model that uses routine MRI to distinguish benign renal lesions from renal cell carcinoma (RCC).

About 65,000 cases of RCC are diagnosed year, accounting for 85% of all renal malignancies. At a Friday session, Thi My Linh Tran, BS, a second-year medical student at Brown University and a researcher in the radiology AI lab at Rhode Island Hospital, discussed the challenge for radiologists in making the distinction between a malignant renal mass and a benign one.

"Because ultrasound, CT and MRI have limited sensitivity and specificity, the current standard of care is to remove any renal mass suspicious of RCC," Tran said. "Of renal masses removed surgically, 20% are benign. There is a clear need for more accurate imaging diagnosis."

For her RSNA 2020 research, "AI Augmentation of Renal Tumor Diagnosis on Routine MR Imaging," Tran was awarded a 2020 RSNA Trainee Research Prize.

Creating the AI Model

In Tran's study, the cohort included 843 patients with confirmed renal lesions collected from two large U.S. academic centers and two large hospitals in China. The study comprised 576 malignant tumors and 267 benign tumors. The cohort was divided into training, validation and test sets in a 7:2:1 ratio.

After manually segmenting T1C and T2WI images on 3D Slicer software, the system was trained using ResNet50 architecture. Clinical variables such as age, gender and tumor volume were fed into a logistic regression model for predicting malignancy. Next, an ensemble model that combined the clinical variable logistic regression, T1C and T2WI models was made using a bagging classifier.

Radiologists Work with AI Assistance

Radiologists with different levels of experience blindly evaluated the test first without and then with AI assistance. "What we found is that the best final model achieved a test accuracy of 82% compared to the radiologists' combined average accuracy of 69%," Tran said.

Tran notes that the test set also outperformed radiologists in terms of sensitivity (82% vs. 74%) and specificity (81% vs. 59%).

"However, when radiologists were assisted by AI, the level of accuracy and sensitivity were higher than what the AI model achieved alone," Tran said.

Specifically, a radiologist working with AI assistance achieved 83% accuracy and 92% sensitivity. Only specificity was lower, at 64% (although still higher than what the radiologist was able to achieve working alone). According to Tran, future work will address the model's low specificity.

"Our study shows how AI assistance improves a radiologist's ability to distinguish benign from malignant renal masses," Tran said. "For the patient, this opens the door to a non-invasive method for diagnosing RCC from benign masses."

Tran's team is currently working to externally validate their model using data from other institutions and are planning a prospective evaluation to deploy the model in a hospital setting.

For More Information:

View the RSNA 2020 session AI Augmentation of Renal Tumor Diagnosis on Routine MR Imaging — SSGU01 at