The prospective study aim to develop a multimodal deep learning model that integrates ultrasound images and cytological whole-slide images for more accurate malignant risk prediction of Bethesda III thyroid nodules.

Objectives: To develop a multimodal deep learning model that integrates ultrasound images and cytological whole-slide images for more accurate malignant risk prediction of Bethesda III thyroid nodules.

Materials and Methods: A ultrasound model, a cytology model, and a fusion model were constructed in this single-center retrospective diagnostic accuracy test. Consecutive patients with Bethesda III thyroid nodules who underwent conventional US examination and fine-needle aspirations were included between January 2016 and December 2024 in Sun Yat-sen Memorial Hospital, Sun Yat-sen University. The reference standard was postoperative histopathology or BRAFV600E mutation. Receiver operating characteristic curve analysis was used to evaluate the diagnostic performance, and decision curve analysis was used to assess the clinical utility.

Inclusion Criteria:

• (a) Patients who underwent fine-needle aspiration of thyroid nodules and received a cytological diagnosis of Bethesda Ⅲ.
• (b) Patients who subsequently underwent thyroid surgery with a definitive histopathological diagnosis of benign or malignant lesions,.
• (c) Availability of at least one ultrasound image acquired prior to surgery.

Exclusion Criteria:

• (a) Patients without confirmation by surgical pathology,.
• (b) Patients without any available or usable ultrasound images.

Patients with a Bethesda Ⅲ thyroid nodule