基于正交分级融合的多模态甲状腺滤泡癌分类算法

hyroid follicular neoplasms encompass two categories: malignant thyroid follicular carcinoma and benign follicular adenoma. Due to their highly similar cytological characteristics, significant challenges remain in clinically differentiating thyroid follicular neoplasms from benign follicular adenomas, while accurate classification is crucial for improving patient treatment and preventing cancer risks. Although existing deep learning models have demonstrated excellent performance in benign-malignant classification of thyroid nodules, their ability to distinguish between thyroid follicular carcinoma and benign follicular adenoma remains limited. This paper proposes a feature-level fusion-based multimodal deep learning model for preoperative non-invasive diagnosis of thyroid follicular carcinoma and benign adenoma. We introduce a novel multimodal fusion strategy incorporating hierarchical fusion and modal orthogonality constraints, facilitating information complementarity and integration across different modalities. The proposed model achieved an accuracy of 92.31% sin thyroid follicular carcinoma classification tasks, outperforming current single-modal and dual-modal fusion approaches, reaching clinically applicable standards.