Malignancy-Aware Dual-Task Heatmap Decoupling for thyroid ultrasound nodule detection and classification

Accurate thyroid nodule detection and benign-malignant classification on ultrasound are critical for early cancer diagnosis. Existing deep-learning detectors encode localization and classification in a shared feature space, causing inter-task feature redundancy and limiting task-specific performance.

We propose Malignancy-Aware Dual-Task Heatmap Decoupling (MADTHD), a CenterNet-based framework that decouples localization and classification through asymmetric dual heads. A Position Head with small receptive fields captures precise nodule contours, while a Malignancy Head with multi-scale dilated convolutions extracts textural and boundary features indicative of benign-malignant status. A feature orthogonality loss is introduced to enforce complementary channel activation between the two heads, reducing inter-task feature redundancy. Systematic ablation studies were conducted on TN5000, and generalization was assessed via zero-shot transfer to the independent TN3K dataset.

MADTHD achieves a mean average precision of 0.7840 on TN5000, outperforming the CenterNet baseline by 2.61 percentage points, with the improvement most pronounced for malignant nodules. Ablation experiments demonstrate that the asymmetric dual-head architecture and the feature orthogonality loss are mutually synergistic. On the independent TN3K dataset, MADTHD maintains its relative standing against the baseline without fine-tuning.

MADTHD provides an interpretable dual-task framework mirroring the clinical localize-then-characterize workflow. The proposed modules are lightweight and backbone-agnostic. Clinical utility remains to be established through prospective validation.