RCAF for patient-level thyroid ultrasound malignancy prediction under leakage-free evaluation and calibration.
Authors
Affiliations (3)
Affiliations (3)
- Department of Computer Science, College of Information Technology, Misr University for Science and Technology (MUST), P.O. Box 77, Giza, Egypt. [email protected].
- College of Information Technology, Misr University for Science and Technology (MUST), P.O. Box 77, Giza, Egypt.
- Department of Computer Science, College of Information Technology, Misr University for Science and Technology (MUST), P.O. Box 77, Giza, Egypt.
Abstract
Accurate differentiation between benign and malignant thyroid nodules on ultrasound remains clinically important, yet interpretation is operator-dependent and subject to inter-observer variability. Recent deep learning studies report strong performance for thyroid ultrasound classification, but many prior approaches remain centred on image-level prediction, with limited emphasis on patient-level baselines, calibration-aware evaluation, and mask-related shortcut analysis. To address these gaps, we present a patient-level thyroid ultrasound malignancy prediction framework centred on a Region-Aware Context-Aware Fusion (RCAF) model evaluated under a strict leakage-free protocol. RCAF combines lesion-focused and context-preserving frame representations through a dual-branch design with gated fusion, followed by attention-based multiple instance learning (AttnMIL) for patient-level aggregation. The framework incorporates development-only probability calibration and threshold selection before single-shot evaluation on an untouched independent test cohort. Experiments on the public ThyroidXL benchmark show that RCAF outperforms stronger fair patient-level baselines, including image-only, transformer-based, and lesion-only comparators. Calibration analysis improves probability reliability, threshold analysis demonstrates stable behaviour under clinically relevant operating conditions, and shortcut sensitivity experiments show that naive mask concatenation produces shortcut-prone gains, whereas RCAF degrades by only 0.001 ROC-AUC under within-patient mask permutation, supporting principled region-aware reasoning. For cross-domain assessment, RCAF was evaluated on TN5000, a Chinese thyroid ultrasound dataset acquired under different imaging conditions; following domain adaptation and 8-view test-time augmentation, the model achieved AUC = 0.914 [0.879-0.948] on a class-balanced validation subset. Overall, RCAF constitutes a strong patient-level thyroid ultrasound classification framework, with encouraging cross-domain adaptability. Broader prospective multi-centre validation remains necessary before clinical deployment.