AI in Medical Imaging Interpretation for Clinical Decision Support

Executive Summary

Detailed Research

Methodology

Evidence is largely derived from diagnostic accuracy studies, reader studies, and meta-analyses evaluating AI algorithms for tasks such as breast-cancer screening, lung-nodule detection, and diabetic-retinopathy grading. These studies compare AI alone, clinicians alone, and AI-assisted clinicians against reference standards.

Outcomes include sensitivity, specificity, area under the ROC curve, and sometimes downstream measures like recall rates, biopsy rates, or time to diagnosis.

Key Studies

Meta-Analyses of Deep Learning in Medical Imaging

• Design: Multiple meta-analyses
• Sample: Thousands of imaging studies
• Findings: Pooled sensitivities and specificities of AI models comparable to specialists for targeted tasks, with AUCs often above 0.90. In combined human+AI conditions, sensitivity typically increases by 5–15 percentage points with modest changes in specificity.
• Clinical Relevance: AI augments rather than replaces expert interpretation

RCTs in Breast-Cancer Screening

• Design: Randomized or paired-reader trials
• Sample: Mammography screening programs
• Findings: AI-supported radiologists detect more cancers at similar recall rates, suggesting potential for improved early detection.
• Clinical Relevance: Promising for population screening

Diabetic Retinopathy Screening

• Design: FDA clearance studies
• Sample: Primary care diabetic populations
• Findings: AI systems demonstrate high sensitivity and specificity compared with ophthalmologists, enabling autonomous or semi-autonomous screening that may expand access.
• Clinical Relevance: Applicable in primary care settings

Clinical Implications

For osteopathic physicians, especially those in primary care, AI-supported imaging interpretation can enhance detection of clinically important findings and support timely referral or intervention, particularly in settings with limited specialist access.

DOs remain responsible for integrating imaging results with structural exam findings, functional status, and patient goals to determine appropriate next steps.

Limitations & Research Gaps

Most evidence focuses on specific imaging tasks; broad, multi-condition AI systems are less well validated. Real-world performance can lag behind test-set results due to differences in patient populations, imaging protocols, and technical integration.

There is limited data on how AI imaging tools affect long-term outcomes, costs, and patient experience when scaled across health systems.

Osteopathic Perspective

Osteopathic medicine values imaging as one element within a broader structural and functional assessment. AI can enhance detection of pathology but cannot replace hands-on examination or appreciation of compensatory patterns.

DOs should use AI-assisted imaging as a supportive tool while maintaining a holistic perspective that integrates imaging with palpatory findings and the patient's overall adaptive response.