An Explainable Hybrid Deep Learning Framework with Adaptive Multi-Level Feature Fusion for Early Thyroid Nodule Detection and Classification Using Ultrasound Images

Abstract

Thyroid ultrasound is widely used to assess thyroid nodules, but reliable interpretation remains challenging because ultrasound images are affected by speckle noise, weak contrast, heterogeneous tissue patterns, and indistinct lesion boundaries. These difficulties become more pronounced for small or irregular nodules. This study presents an Explainable Hybrid Deep Learning Framework with Adaptive Multi-Level Feature Fusion (AMFF) for thyroid nodule analysis. The framework combines a CNN branch with a Transformer/Mamba branch to learn complementary local and contextual representations. AMFF is introduced to weight information from different feature levels rather than relying on a fixed feature-combination scheme. A boundary-aware segmentation component is used to refine lesion delineation, and segmentation and malignancy classification are addressed within a multi-task architecture. Grad-CAM++ is incorporated to provide visual evidence associated with the classification decision. The resulting framework is designed as an integrated computer-aided analysis pipeline for thyroid ultrasound images.