Noting overlapping imaging features that can make it challenging to differentiate small renal masses (SRMs) on computed tomography (CT) scans, the authors of a new study suggest that a deep learning algorithm offers comparable detection to that of urological radiologists and superior performance in comparison to non-urological radiologists.
For the retrospective study, recently published in Radiology, researchers reviewed CT scans for 1,703 patients (mean age of 56) who had single renal masses. After development of the deep learning algorithm in training and internal test sets, the study authors assessed the effectiveness of the algorithm for detecting benign SRMs < 3 cm and < 1 cm in multicenter external testing and prospective testing.
For benign SRMs < 3 cm, the multicenter external testing revealed an 80 percent area under the curve (AUC) for the deep learning algorithm, which was less than the AUC for urological radiologists (84 percent) but six percent higher than the non-urological radiologist average (74 percent) and 14 percent higher than the urologist average (66 percent).
The researchers also noted the algorithm had lower sensitivity in comparison to urological radiologists (48 percent vs. 59 percent) but the sensitivity rate was 21 and 22 percent higher, respectively, than that of non-urological radiologists (27 percent) and urologists (26 percent).
“Translating the artificial intelligence research into the optimization of the clinical workflow is the ultimate goal. We found that general radiologists and urologists are less capable of identifying benign SRMs. Our DL algorithm could assist less experienced physicians when expert radiologists are absent or unavailable in resource-poor hospitals,” wrote lead study author Chenchen Dai, M.D., who is affiliated with the Department of Radiology at the Zhongshan Hospital at Fudan University, and the Shanghai Institute of Medical Imaging in Shanghai, China, and colleagues.
While the deep learning algorithm had comparable AUC to urological radiologists in prospective testing (90 percent vs. 91 percent) for sub-centimeter renal masses (< 1) on CT, the researchers noted the algorithm had a 22 percent lower sensitivity rate in external multicenter testing (26 percent vs. 58 percent). However, the sensitivity of the algorithm for sub-centimeter renal masses was more than double that of non-urological radiologists (11 percent), according to the external multicenter data.
The researchers conceded that limited resolution and scarce pathology types may have contributed to the increasing misclassification rate they saw with the algorithm for the sub-centimeter lesions.
Three Key Takeaways
- Deep learning algorithm performance. The deep learning algorithm showed promising performance in detecting small renal masses (SRMs) on CT scans, demonstrating comparable detection rates to urological radiologists and superior performance compared to non-urological radiologists. This suggests that the algorithm could be a valuable tool in assisting less experienced physicians and optimizing clinical workflow, especially in resource-poor hospitals.
- Detection of benign SRMs. The algorithm exhibited an 80 percent area under the curve (AUC) for detecting benign SRMs smaller than 3 cm in multicenter external testing. Although the sensitivity of the algorithm was lower than that of urological radiologists, it outperformed both non-urological radiologists and urologists, suggesting its potential in accurately identifying benign lesions and reducing unnecessary surgeries.
- Prospective testing and potential impact. In prospective testing for sub-centimeter renal masses (< 1 cm) on CT, the deep learning algorithm demonstrated comparable AUC to urological radiologists. While it had a lower sensitivity rate in external multicenter testing, it still showed substantial improvement over non-urological radiologists. The study authors suggested that limited resolution and scarce pathology types may have contributed to the reduced sensitivity of the algorithm for sub-centimeter renal masses.
However, the study authors suggested that the overall potential of the algorithm could have an impact in triaging cases involving renal masses of < 3 cm.
“The (deep learning) algorithm could act as the primary readers of kidney tumor CT images and reduce the workload for radiologists. If the DL result indicates that an SRM is benign, it prompts urological radiologists to re-examine and analyze the case more carefully. When an obvious discrepancy arises, it is necessary to perform active surveillance or a biopsy to confirm the diagnosis, thereby reducing unnecessary surgeries to some extent,” posited Dai and colleagues.
(Editor’s note: For related content, see “Can a CT-Based Radiomics Model Enhance Risk Stratification for Clear Cell Renal Cell Carcinoma?,” “Study: PSMA PET/CT Identifies 18 Percent More Metastatic Renal Cancers than Conventional Imaging” and “Emerging PET/CT Agent May Enhance Diagnosis for Smaller Lesions of Clear Cell Renal Cell Carcinoma.”)
In regard to study limitations, the authors noted the deep learning algorithm was entirely based on data from surgical patients and primarily developed with CT slice thickness of 5 mm. The researchers acknowledged that employing thinner CT slices may have improved the algorithm’s segmentation and accuracy with classification of masses.
