Asan Medical Center AI Raises Rare Kidney Cancer Diagnosis Accuracy

Using deep learning artificial intelligence to detect a rare kidney cancer raises diagnostic accuracy from 60.0% to 73.3%.

Key Clinical Takeaways:

• Diagnostic accuracy improved by 13.3 percentage points when physicians used the AI screening tool.
• Physicians who heavily relied on the AI caught correct diagnoses more often but also accepted incorrect AI outputs more frequently.
• AI utilization tendencies showed no clear link to a doctor’s clinical experience or medical specialty.

The research evaluated detection rates for fumarate hydratase-deficient renal cell carcinoma, known as FHdRCC. A team led by professor Nam-guk Kim in the department of convergence medicine and professor Young-mi Cho in the department of pathology at Asan Medical Center developed the deep learning tool and tested it with 21 pathologists.

FHdRCC accounts for less than 0.4% of all kidney cancers. The disease progresses rapidly, carries a poor prognosis, and mimics other renal malignancies during standard pathological review.

Researchers analyzed 39 patients diagnosed with the rare cancer. Their average age was 47.5 years, with 67% diagnosed at advanced stages and 39% showing lymph node metastasis. Only 36% of these cases were initially suspected during first-line evaluation.

The team trained the screening AI on 335 digital pathology tissue images. Internal evaluation yielded an area under the curve of 0.954, while external evaluations reached 0.820 and 0.867.

Experiments compared readings from 21 medical professionals, including seven urological pathologists, seven pathologists from other subspecialties, and seven resident physicians. Each clinician reviewed 30 difficult tissue slides twice—once without AI and once with AI assistance—separated by a two-week washout period.

Reading times did not increase during the experiment. However, clinicians split into two behavioral groups based on exploratory cluster analysis: 14 physicians actively accepted AI judgments, while seven maintained a cautious stance.

Active adopters capitalized more effectively on correct AI outputs, showing a median utilization score of 0.732 compared to 0.444 for the cautious group. At the same time, active adopters also accepted incorrect AI outputs more readily, with an over-reliance median score of 0.800 versus 0.600 for cautious readers.

The experimental set contained an equal 50-50 split of rare cancer cases and other renal tumors, a distribution that does not reflect real-world clinical prevalence. The AI generated incorrect outputs in only five of the test cases.

“This study is meaningful because it analyzed the implications of developing AI for rare pathology diagnostics and applying it in actual clinical settings,” Kim said, adding that the findings show how “the ability to effectively utilize AI is an independent competency from existing diagnostic skill or experience.”

“Rare kidney cancers are difficult to diagnose even for experienced pathologists, and if the disease itself is not suspected, additional tests required for accurate diagnosis may not be performed,” Cho said, noting that “this study showed the possibility that AI can provide practical help in raising early suspicion for rare cancers that are easy to miss and selecting patients who need additional tests.”

Disclaimer: The information provided in this article is for educational and scientific communication purposes only and does not constitute medical advice. Always consult with a qualified healthcare provider regarding any medical condition, diagnosis, or treatment plan.