Deep Learning-Based Recognition of Different Thyroid Cancer Categories Using Whole Frozen-Slide Images

Abstract

Introduction: The pathological rare category of thyroid is a type of lesion with a low incidence rate and is easily misdiagnosed in clinical practice, which directly affects a patient's treatment decision. However, it has not been adequately investigated to recognize the rare, benign, and malignant categories of thyroid using the deep learning method and recommend the rare to pathologists. Methods: We present an empirical decision tree based on the binary classification results of the patch-based UNet model to predict rare categories and recommend annotated lesion areas to be rereviewed by pathologists. Results: Applying this framework to 1,374 whole-slide images (WSIs) of frozen sections from thyroid lesions, we obtained an area under a curve of 0.946 and 0.986 for the test datasets with and without WSIs, respectively, of rare types. However, the recognition error rate for the rare categories was significantly higher than that for the benign and malignant categories (p < 0.00001). For rare WSIs, the addition of the empirical decision tree obtained a recall rate and precision of 0.882 and 0.498, respectively; the rare types (only 33.4% of all WSIs) were further recommended to be rereviewed by pathologists. Additionally, we demonstrated that the performance of our framework was comparable to that of pathologists in clinical practice for the predicted benign and malignant sections. Conclusion: Our study provides a baseline for the recommendation of the uncertain predicted rare category to pathologists, offering potential feasibility for the improvement of pathologists' work efficiency.

Keywords: WSI; deep learning model; pathology; rare category; thyroid cancer.

FIGURE 1

An overview of the proposed WSI diagnostic framework presented in this study. (A) The WSI slide with the region of interest (green line) and carcinoma region (blue line). (B) The process of patch-based UNet model training. (C) The process of patch-based UNet inference. (D) The WSI heatmap. (E) A random forest was selected for the WSI-based classification task. (F) The proposed triple classification model.

FIGURE 2

(A) AUC for the different test datasets. The Test 1 dataset contained common benign and malignant pathological subtypes (PTC and NG). Test 2 contained not only the common PTC and NG types included in Test 1 but also two intermediate types (TAL and TFCN) and two other rare types (other TC and BTL). (B) The confusion matrix for the benign, malignant, and intermediate subtypes.

FIGURE 3

Misclassification examples of selected slides in the Test 1 dataset. Examples of true positive, false negative, false positive, and other subtype slides are represented. Our model performed efficiently in tumor carcinomas for true positives (PTC), but the false positives (NG) were mainly caused by the fibrotic tissue in both the normal and carcinoma regions. Fibrotic tissue sometimes has a larger area or diameter than certain carcinoma regions for false negatives (PTC). Because fibrotic tissue is quite common in thyroid WSIs, the TAL, TFCN, and other BTL slides outside our training slides showed clear differences in structural features from PTC and NG slides, resulting in a heatmap far from the ground truth.

FIGURE 4

Decision tree for distinguishing the rare from the common benign and malignant categories.