Deep Learning for Ultrasound Classification to Identify Noninvasive Follicular Thyroid Neoplasms with Papillary-Like...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Noninvasive follicular thyroid neoplasm with papillary-like nuclear features (NIFTP) shares clinical and demographic overlap with other encapsulated thyroid tumors.
• Accurate preoperative differentiation of these tumors is challenging due to overlapping features.

Purpose of the Study:

• To assess the feasibility of using deep learning (DL) models on ultrasound images for classifying and identifying NIFTPs.
• To evaluate the performance of DL models in distinguishing NIFTPs from other encapsulated follicular thyroid lesions.

Main Methods:

• ResNet50 and EfficientNet_B0 deep learning models were employed for feature extraction from preoperative ultrasound images.
• A dataset of 279 encapsulated follicular thyroid tumors (including follicular adenomas, follicular thyroid cancers, NIFTPs, and invasive encapsulated follicular variant papillary thyroid cancers) was utilized.
• Models were trained and internally validated, followed by external validation on a separate prediction cohort.

Main Results:

• The EfficientNet model achieved higher internal validation accuracy (0.95) compared to ResNet50 (0.88).
• Both models demonstrated modest performance on the external prediction cohort, with an accuracy of 0.77.
• Gradient-weighted class activation mapping (Grad-CAM) revealed that models focused on nodule parenchyma for classification.

Conclusions:

• Deep learning models show potential for preoperative differentiation of encapsulated follicular thyroid tumors, including NIFTPs.
• Despite overlapping features, AI-driven analysis of ultrasound images offers promising accuracy, sensitivity, and specificity for improved diagnostic capabilities.