TAZD-Net: zone-aware differentiable morphology for joint thyroid nodule segmentation and malignancy classification

Thyroid ultrasound is the primary imaging modality for thyroid nodule assessment, yet reliable malignancy prediction remains challenging because diagnostic reasoning depends not only on lesion appearance but also on morphology, boundary characteristics, peri-lesional transitions, and surrounding tissue context. Existing multi-task approaches often underexploit this spatial organization and rarely incorporate explicit morphology as a differentiable component of end-to-end optimization. We present TAZD-Net, a segmentation-conditioned multi-task framework for joint thyroid nodule delineation and binary malignancy classification from B-mode ultrasound images. An EfficientNet-B4 encoder and feature-pyramid decoder generate a soft lesion representation that guides differentiable decomposition of the nodule environment into lesion, peri-lesional rim, and surrounding-context regions. Region-specific features are combined with a differentiable morphology representation derived directly from the soft segmentation and interact through a cross-region Transformer with sample-specific multiplicative gating for malignancy prediction. On the patient-disjoint ThyroidXL held-out test set of 2094 images from 739 patients, TAZD-Net achieved a patient-mean Dice coefficient of 0.8825 (95% CI, 0.8759–0.8883), image-level AUROC of 0.9325 (95% CI, 0.9149–0.9481), and patient-level AUROC of 0.9576 (95% CI, 0.9441–0.9700), with a patient-level F1-score of 0.8554 (95% CI, 0.8262–0.8842). Alternative aggregation strategies produced distinct ranking and threshold-dependent behavior, while component-removal experiments showed generally modest effects on headline AUROC and Dice. TAZD-Net therefore integrates lesion delineation, anatomically structured regional reasoning, differentiable morphology, and patient-level malignancy inference within a unified framework.