ThyroNet-XAI: An Explainable Multi-Modal Deep Learning System for Early Detection and Classification of Thyroid Diseases Using Kolmogorov-Arnold Networks

Abstract

Endocrinology still faces difficulty in detecting and classifying thyroid disorders ranging from subtle hidden dysfunctions to obvious thyroiditis and nodules. Single-modality diagnostic tools often lack sensitivity when confronted with complex hormone interactions or overlapping systemic conditions. To address these shortcomings, ThyroNet-XAI is proposed as a solution. This novel explainable multi-modal deep learning framework synergistically integrates structured clinical records, endocrinological biomarkers, and multi-parametric medical imaging (US, MRI, and SPECT). In our approach, we replace standard Multi-Layer Perceptions (MLPs) with Kolmogorov-Arnold Networks (KANs), which use adaptive B-spline activation functions along network edges to model complex biochemical dependencies. To ensure interpretability, we use a dynamic multi-head cross-attention mechanism and Shapley Additive explanations (SHAP). As part of the validation process, we used the TNS000 Nodule Dataset and the UCI Thyroid Dataset to validate ThyroNet-XAI. We find that ThyroNet-XAI performs better than state-of-the-art multi-modal baselines, providing clinically interpreted feature rankings while achieving 98.65% classification accuracy.