Using deep convolutional neural networks for multi-classification of thyroid tumor by histopathology: a large-scale...

Area of Science:

• Pathology
• Computer Science
• Oncology

Background:

• Histopathological slide analysis is crucial for thyroid nodule diagnosis.
• Interobserver variability can impact diagnostic accuracy.
• Deep convolutional neural networks (DCNNs) offer potential for automated image analysis.

Purpose of the Study:

• To evaluate the diagnostic efficiency of DCNNs in classifying thyroid nodules from histopathological images.
• To assess the potential of DCNNs to improve interobserver agreement in thyroid cancer diagnosis.

Main Methods:

• Trained Inception-ResNet-v2 and VGG-19 models on 11,715 fragmented histological images from 806 patients.
• Classified seven types: normal tissue, adenoma, nodular goiter, papillary thyroid carcinoma (PTC), follicular thyroid carcinoma (FTC), medullary thyroid carcinoma (MTC), and anaplastic thyroid carcinoma (ATC).
• Analyzed diagnostic efficiencies and misdiagnoses for each pathology type.

Main Results:

• VGG-19 demonstrated higher average diagnostic accuracy (97.34%) compared to Inception-ResNet-v2 (94.42%).
• VGG-19 achieved excellent diagnostic efficiencies for all malignant thyroid cancer types (e.g., 100% for MTC and nodular goiter).
• Normal tissue and adenoma were identified as the most challenging types for DCNN classification.

Conclusions:

• DCNN models, particularly VGG-19, achieve high accuracy in differentiating thyroid tumors histopathologically.
• DCNNs show potential to enhance diagnostic efficiency and consistency in thyroid pathology.
• Further development could lead to DCNNs assisting pathologists in thyroid disease diagnosis.