Deep Learning for Pathology: Optimizing CNN Architectures for Papillary Thyroid Carcinoma Classification

Abstract

Papillary thyroid carcinoma (PTC) is the most common type of thyroid cancer and has a high cure rate if detected in the early stage. Pathologists diagnose PTC by histology or cytology, paying special attention to tumour nuclear features. Despite being highly accurate in the hands of skilled observers, this approach is not entirely reliable, as some cases may lack classical features. The remarkable advances in artificial intelligence (AI) technology offer the possibility of assisting pathologists in the diagnosis of cancers such as PTC. In this study, convolutional neural networks (CNNs) were applied to classify histopathological images from 60 PTC cases, using normal adjacent thyroid tissue as control. The dataset was restructured at the patient level to prevent data leakage by ensuring that images from the same patient were assigned to a single subset. Five CNN architectures (VGG16, VGG19, ResNet50V2, DenseNet121, and EfficientNetB0) were evaluated using four architectural variations. Model selection was performed using validation data, followed by optimization of class-weight and threshold strategies. DenseNet121 achieved the best performance, with an average accuracy of 0.9883 and sensitivity of 0.9871. These results demonstrate that optimized CNN models can effectively identify PTC from histopathological images. Patient-level splitting provides a more realistic evaluation of model performance. The proposed approach shows strong potential as a decision-support tool, although further validation on multi-institutional datasets is required.