Adaptive Bandelet Transform and Transfer Learning for Geometry‑Aware Thyroid Cancer Ultrasound Classification

Accurate classification of thyroid nodules in ultrasound remains challenging due to limited labeled data and the weak ability of conventional feature representations to capture complex, multi-directional textures. To address these issues, we propose a geometry-aware framework that integrates the adaptive Bandelet Transform (BT) with transfer learning (TL) for benign–malignant thyroid nodule classification. The method first applies BT to enhance directional and structural encoding of ultrasound images through quadtree-driven geometric adaptation, then mitigates class imbalance using SMOTE and expands data diversity via targeted augmentation. The resulting images are classified using several ImageNet-pretrained architectures, with VGG19 providing the most consistent performance. Experiments on the publicly available DDTI dataset show that BT-based preprocessing improves performance over classical wavelet representations across multiple quadtree thresholds, with the best results achieved at T=30. Under this setting, the proposed BT+TL(VGG19) model attains 98.91% accuracy, 98.11% sensitivity, 97.31% specificity, and a 98.89% F1-score, outperforming comparable approaches reported in the literature. These findings suggest that coupling geometry-adaptive transforms with modern TL backbones can provide robust, data-efficient ultrasound classification. Future work will focus on validating generalizability across larger multi-centre datasets and exploring transformer-based classifiers.