Oral squamous cell carcinoma diagnosis in digitized histological images using convolutional neural network

Diagnostic methods of oral squamous cell carcinoma (SCC) using artificial intelligence (AI) and digital-histopathologic images have been developed. However, previous AI training methods have focused on the cellular atypia given by the training of high-magnification images, and little attention has been paid to structural atypia provided by low-power wide fields. Since oral SCC has histopathologic types with bland cytology, both cellular atypia and structural atypia must be considered as histopathologic features. This study aimed to investigate AI ability to judge oral SCC in a novel training method considering cellular and structural atypia and their suitability.

We examined digitized histological whole-slide images from 90 randomly selected patients with tongue SCC who attended a dental hospital. Image patches of 1000 × 1000 pixels were cut from whole-slide images at 0.3125-, 1.25-, 5-, and 20-fold magnification, and 90,059 image patches were used for training and evaluation. These image patches were resized into 224 × 224, 384 × 384, 512 × 512, and 768 × 768 pixels, and the differences in input size were analyzed. EfficientNet B0 was utilized as the convolutional neural network model. Gradient-weighted class activation mapping (Grad-CAM) was used to elucidate its validity.

The proposed method achieved a peak accuracy of 99.65% with an input size of 512 × 512 pixels. Grad-CAM suggested that AI focused on both cellular and structural atypia of SCC, and tended to focus on the region surrounding the basal layer.

Training AI regarding both cellular and structural atypia using various magnification images simultaneously may be suitable for the diagnosis of oral SCC.