Chest CT bilateral-asymmetry AI shows modest specificity but low sensitivity for breast cancer classification

Bilateral-asymmetry multiple-instance learning for breast cancer on preselected chest CT slices achieved sensitivity 0.286 (95% CI 0.160-0.415), specificity 0.825, and ROC AUC 0.547 in nested validation—feasibility only, not detection of unsuspected cancer.

• Retrospective single-center case-control study of 89 patients (49 breast cancer, 40 non-malignant) with 304 preselected axial chest CT images arranged as 304 bilateral crop pairs.
• Primary nested patient-level five-fold cross-validation yielded accuracy 0.528, balanced accuracy 0.555, sensitivity 0.286, specificity 0.825, and ROC AUC 0.547.
• Secondary non-nested analysis gave higher exploratory estimates but used held-out folds for early stopping; findings support feasibility on preselected images, not detection across complete CT examinations.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.