search

NEWS AI Improves Diagnostic Accuracy for Rare Kidney Cancer, but Its Effectiveness Depends on Physician Acceptance 2026.11.19

Asan Medical Center Develops AI to Aid Rare Kidney Cancer Diagnosis, Offering Potential to Identify Patients Who Need Further Testing

 

Diagnostic accuracy among 21 physicians improved from 60.0% to 73.3%. Greater physician acceptance of AI enhanced diagnostic performance but also increased the risk of accepting incorrect AI recommendations.

 

▲ (From left) Professors Namkug Kim from the Department of Convergence Medicine and Yong Mee Cho from the Department of Pathology at Asan Medical Center

 

A recent study has found that artificial intelligence (AI) can improve diagnostic accuracy in pathology, which involves examining tissue or cell samples collected from the human body to diagnose diseases, including rare cancers. However, the ability to use AI effectively is a distinct skill that is independent of a physician’s diagnostic experience or expertise. In particular, excessive reliance on AI when diagnosing rare diseases may lead to incorrect diagnoses, highlighting the need for caution.

 

A research team led by Professors Namkug Kim from the Department of Convergence Medicine and Yong Mee Cho from the Department of Pathology at Asan Medical Center developed a deep learning AI model to screen for fumarate hydratase-deficient renal cell carcinoma (FHdRCC), a rare type of kidney cancer. In an evaluation involving 21 physicians, the team found that diagnostic accuracy improved from 60.0% to 73.3% with AI assistance.

 

FHdRCC accounts for less than 0.4% of all kidney cancers and is characterized by rapid progression and a poor prognosis. Because its tissue morphology is similar to that of other kidney cancers, it can be difficult to distinguish, and additional tests, including genetic testing, are required for a definitive diagnosis. However, if the disease is not initially suspected, these additional tests may not be performed.

 

This cancer was first included in the World Health Organization (WHO) classification in 2016. Since most pathologists have not encountered it during their training, it is an area of diagnosis in which AI assistance could offer significant clinical benefits.

 

Among the 39 patients included in the study, the mean age was 47.5 years, and 67% were diagnosed at an advanced stage. Lymph node metastasis was found in 39% of patients, and the five-year survival rate was 61.8%. Only 36% of cases were initially suspected to be FHdRCC.

 

Professor Namkug Kim’s team developed and evaluated the AI model using 335 digital pathology images. The model achieved an area under the curve (AUC) of 0.954 in internal validation and 0.820 and 0.867 in two external validation datasets. AUC measures a model’s ability to distinguish between the presence and absence of a disease, with values closer to 1 indicating better performance.

 

The research team also asked 21 physicians, including pathologists and pathology residents, to interpret 30 challenging kidney cancer tissue samples, both with and without AI assistance, and compared their diagnostic results.

 

With AI assistance, diagnostic accuracy significantly improved from 60.0% to 73.3%, while precision increased from 66.7% to 77.8%. Interobserver agreement among pathologists also rose from 0.311 to 0.482, with no significant increase in interpretation time.

 

Notably, the team identified two distinct patterns of physician behavior in using AI: an “accepting” approach, in which physicians readily incorporated AI recommendations, and a “cautious” approach, in which they carefully reviewed AI suggestions before making a diagnosis.

 

Physicians in the accepting group showed significantly greater improvements in diagnostic accuracy, as they were more likely to follow correct AI recommendations. However, they also had a higher rate of accepting incorrect AI recommendations, at 0.800, compared with 0.600 in the cautious group. Physicians in the cautious group were less influenced by AI errors, but the improvement in diagnostic accuracy was more limited.

 

Physicians’ AI adoption patterns were not associated with their clinical experience or subspecialty. The research team noted that, particularly when diagnosing rare diseases through pathology, diagnostic expertise and the ability to use AI effectively may represent two distinct skill sets. They emphasized the need for AI training programs that take individual physicians’ adoption patterns into account.

 

“This study is significant in that we developed an AI model for rare pathology diagnoses and examined the implications of using it in actual clinical practice. Our findings show that the ability to use AI effectively is an independent skill, separate from existing diagnostic expertise or clinical experience,” said Professor Namkug Kim of the Department of Convergence Medicine at Asan Medical Center. “Going forward, we need training programs and collaboration strategies tailored to each physician’s individual approach to using AI.”

 

“Rare kidney cancers can be difficult to diagnose even for experienced pathologists. If the disease is not initially suspected, additional tests required for an accurate diagnosis may not be performed,” said Professor Yong Mee Cho of the Department of Pathology at Asan Medical Center. “This study demonstrates the potential of AI to help physicians identify rare cancers that might otherwise be overlooked and select patients who need further testing.”

 

The study was recently published in npj Digital Medicine, an international journal with an impact factor of 18.0.

Back

ASAN MEDICAL CENTER NEWSROOM

PRIVACY POLICY

GO