Deep learning for multitask prediction on thyroid nodule frozen sections.

Area of Science:

• Pathology
• Oncology
• Artificial Intelligence

Background:

• Intraoperative frozen sections are crucial for thyroid nodule diagnosis but face challenges like misdiagnosis and pathologist shortages.
• Deep learning and radiomics show potential for improving thyroid nodule diagnostics.
• Integration of deep learning in intraoperative thyroid nodule analysis is underexplored.

Purpose of the Study:

• Develop deep learning models for intraoperative pathological diagnosis of thyroid nodules.
• Classify thyroid nodules as benign or malignant.
• Predict BRAF V600E gene mutation and lymph node metastasis.

Main Methods:

• Analyzed 436 Whole-Slide Images (WSIs) of thyroid frozen sections using deep learning.
• Employed image preprocessing, feature extraction, and classifier training.
• Utilized Patch Likelihood Histogram (PLH) and Bag of Words (BoW) for patch-to-WSI feature aggregation.

Main Results:

• InceptionV3 achieved an AUC of 0.998 for benign/malignant classification, outperforming supervised methods with weakly supervised strategies.
• ResNet50 predicted BRAF V600E mutation with WSI-level accuracy of 94.4%.
• A ViT-based model achieved 76% accuracy for lymph node metastasis prediction.

Conclusions:

• Deep learning models effectively aid in classifying thyroid frozen sections and predicting BRAF mutations and lymph node metastasis.
• Weakly supervised strategies are effective for thyroid lesion frozen sections, potentially reducing the need for extensive pathologist annotations.
• These models offer a promising approach to enhance intraoperative diagnosis of thyroid nodules.