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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Thyroid nodule classification in ultrasound is difficult due to limited data and texture complexity.
• Conventional methods struggle to capture intricate, multi-directional textures in ultrasound images.

Purpose of the Study:

• To improve data-efficient thyroid nodule classification.
• To enhance feature representation and generalization using Bandelet Transform (BT) and transfer learning (TL).

Main Methods:

• Applied geometry-adaptive Bandelet Transform (BT) for enhanced directional and structural encoding.
• Mitigated class imbalance with SMOTE and increased data diversity via augmentation.
• Classified features using ImageNet-pretrained architectures, with VGG19 showing best performance.

Main Results:

• BT preprocessing improved performance over wavelet representations across various thresholds.
• The BT+TL (VGG19) model achieved 98.91% accuracy, 98.11% sensitivity, 97.31% specificity, and 98.89% F1-score.
• Outperformed comparable methods on the DDTI dataset.

Conclusions:

• Combining geometry-adaptive transforms with TL backbones offers a robust, data-efficient strategy for thyroid nodule classification.
• This approach is particularly effective with limited annotations and complex textures.
• The project and code are publicly available.