AI 診斷系統在三大主要疾病類別中現已比專科醫師更準確

A landmark study published this week in the New England Journal of Medicine has found that AI diagnostic systems outperform board-certified specialists in detecting early-stage disease in three major categories: lung cancer on CT scans, diabetic retinopathy on retinal photographs, and melanoma on dermoscopy images. The study, conducted across 47 clinical sites in 12 countries with over 180,000 patient cases, is the largest and most rigorous evaluation of AI diagnostic performance ever conducted.

The results are striking. For early-stage lung cancer detection, the AI system achieved a sensitivity of 94.3% at a specificity of 91.7% — compared to a mean sensitivity of 87.1% and specificity of 85.3% for the 312 radiologists who participated in the study. For diabetic retinopathy, the AI system's performance was even more pronounced, with a sensitivity of 97.8% compared to 89.4% for ophthalmologists. For melanoma, the gap was smaller but still statistically significant: 91.2% versus 86.7%.

The study's authors are careful to contextualize these findings. The AI systems were evaluated under conditions that favor AI performance: standardized image acquisition protocols, well-curated datasets, and evaluation on the specific task of binary classification. Real-world clinical diagnosis involves far more complexity — integrating imaging findings with patient history, physical examination, laboratory results, and clinical judgment about the appropriate next steps. None of these contextual factors were captured in the study.

"These results are impressive and important, but they do not tell us that AI should replace radiologists or ophthalmologists. They tell us that AI can be a very powerful tool in the hands of those specialists — one that could help catch the cases that would otherwise be missed."

— Lead author, NEJM study

The clinical implications are significant regardless of how one interprets the comparison with specialists. Early detection is the single most important factor in cancer outcomes: five-year survival rates for early-stage lung cancer exceed 90%, compared to less than 10% for late-stage disease. If AI systems can identify cases that would otherwise be missed or delayed, the impact on patient outcomes could be substantial.

The study also documents significant variation in AI performance across demographic groups. The system's sensitivity was 3-5 percentage points lower for patients from underrepresented racial and ethnic groups, a disparity that the authors attribute to underrepresentation of these groups in the training data. This finding reinforces concerns that have been raised about AI diagnostic systems perpetuating or amplifying existing health disparities if deployed without careful attention to training data composition and ongoing performance monitoring.

Regulatory responses to the study are already emerging. The FDA has indicated that it will fast-track review of AI diagnostic systems that demonstrate performance meeting the thresholds established in the study. The European Medicines Agency has issued guidance suggesting that AI diagnostic systems in the three studied categories may be eligible for conditional approval pending real-world performance monitoring. Several major health systems have announced plans to begin pilot deployments of AI diagnostic tools as adjuncts to specialist review.

The economic implications are also significant. Specialist physician shortages are a global problem, particularly in radiology and ophthalmology. If AI systems can perform initial screening at specialist-level accuracy, they could dramatically expand access to diagnostic services in underserved regions and reduce the burden on specialist physicians, allowing them to focus on the most complex cases. The study estimates that widespread deployment of AI diagnostic tools could prevent approximately 180,000 preventable cancer deaths annually in the United States alone.