Symmetry-Consistent Attention Supervision for Explainable Multi-label Temporal-Bone CT Assessment

The purpose of this study is to develop an explainable deep learning framework for region-wise, multi-label assessment of inner- and middle-ear status on temporal-bone computed tomography (CT). The framework used a ResNet-18 encoder with class-specific attention pooling to predict eight labels representing normal and congenital malformation states of the left and right inner and middle ears. Sparse polygon annotations supervised attention using masked binary cross-entropy, Dice-overlap, and outside-mask penalties. A left–right equivariant consistency regularizer enforced bilateral symmetry under horizontal flipping. The dataset comprised 2399 axial and coronal CT images from 87 patients. Development used fourfold patient-level cross-validation on 35 patients (1018 images), followed by evaluation on a held-out internal test set of 52 patients (1381 images). In cross-validation, the full model achieved a macro-AUC of 0.765 and a macro-F1 of 0.627. On the held-out internal test set, image-level evaluation yielded a macro-AUC of 0.688 and a macro-F1 of 0.572, representing the highest mean performance among the evaluated CNN, vision transformer, and MedViT baselines. Patient-level aggregation further yielded a macro-AUC of 0.716 and a macro-F1 of 0.621. The proposed attention maps were also more side-specific and anatomically concentrated than baseline heatmaps. Attention supervision and bilateral equivariance may improve interpretable temporal-bone CT assessment. Larger external studies and patient-level evaluation are required before clinical deployment.