Bilateral-Asymmetry-Aware Multiple-Instance Learning for Breast Cancer Classification on Chest CT.

Incidental breast abnormalities may be visible on chest computed tomography (CT), although breast tissue is frequently outside the primary diagnostic focus of these examinations. This proof-of-concept study developed and internally evaluated a bilateral-asymmetry-aware multiple-instance learning framework (BAA-MIL) for case-level discrimination between breast cancer and non-malignant control examinations using preselected breast-containing axial chest CT images. The retrospective single-center case-control dataset included 89 CT examinations from 89 unique patients: 49 histopathologically confirmed breast cancer cases and 40 non-malignant control cases established according to predefined case-level reference criteria. The dataset comprised 304 selected axial CT images, yielding 608 unilateral breast-region crops arranged as 304 bilateral crop pairs. Paired breast crops were processed by a shared convolutional encoder, combined through an explicit side-symmetric asymmetry descriptor, and aggregated across axial images using gated attention-based multiple-instance learning. Performance was assessed using exploratory five-fold stratified patient-level cross-validation and a nested patient-level validation analysis. Under the primary nested patient-level validation protocol, the compact-encoder BAA-MIL implementation achieved an accuracy of 0.528 (95% CI, 0.427-0.629), balanced accuracy of 0.555 (95% CI, 0.465-0.637), sensitivity of 0.286 (95% CI, 0.160-0.415), specificity of 0.825 (95% CI, 0.694-0.930), and ROC AUC of 0.547 (95% CI, 0.420-0.665). The original non-nested analysis yielded higher exploratory estimates, but this analysis used the held-out folds for early stopping and checkpoint selection and is therefore reported only as secondary exploratory analysis. These findings support technical feasibility for a constrained case-level classification task using preselected chest CT images, but do not demonstrate detection of unsuspected incidental breast cancer across complete CT examinations. Performance estimates remain uncertain, and external validation using complete volumetric CT data is required before clinical use.