Can artificial intelligence support less experienced radiologists in interpreting indeterminate mammographic screenings?

Abstract

Artificial intelligence (AI) has emerged as a promising tool for mammography interpretation that could potentially support radiologists in effectively classifying indeterminate cases. This study analyzed data from 169 less experienced radiologists with less than a year of experience interpreting mammograms, who collectively interpreted 22,200 mammogram cases. These readers categorized cases using the Tabar Grading system, with grade 3 representing indeterminate. Radiologists interpreting >20 cases per week classified 24.8% of mammograms as indeterminate. For those interpreting 20-60 cases per week, this increased slightly to 26.0%. The highest rate was among radiologists handling 61-100 cases per week, with 29.4% classified as indeterminate. Among those reading 100+ cases per week, the percentage dropped slightly to 24.9%. The Globally-aware Multiple Instance Classifier (GMIC) AI model was fine-tuned using a locally acquired dataset to achieve an area under the receiver operating characteristic curve of 0.85+. The images classified as indeterminate by each reader were then fed into GMIC. We explored GMIC’s malignancy probabilities to assess the percentage of users with improved specificity without losing sensitivity. We considered lax, moderate, and strict thresholds. Overall, at the best threshold (lax), we observed an increase in specificity for 80%+ of readers without a loss in sensitivity. With moderate or strict thresholds, the effectiveness of AI in enhancing specificity without a loss in sensitivity diminished. This finding suggests that while AI is a valuable tool for improving diagnostic accuracy in indeterminate cases, optimizing AI settings is essential to maximize its benefits for less experienced readers.

Ähnliche Arbeiten