Using Deep Learning Techniques for an Early Detection of Oral Epithelial Dysplasia

Abstract

Oral cancer is ranked as the sixth most common type of cancer worldwide, with 90% of cases being oral squamous cell carcinoma (OSCC). OSCC has a high mortality rate, and early diagnosis can increase the survival rate. About 80% of OSCC is developed from Oral Epithelial Dysplasia (OED); thus, OED detection is critical to diagnose OSCC at the early stage. Traditionally, the OED is defined by sixteen criteria, including architectural and cytological features, under the microscope by expert oral pathologists. This manual detection is a time-consuming and tedious task, and thus, there is a need for precise automated diagnostic and classification techniques. However, disengaging a Computer Aided Diagnosis (CAD) for OED is challenging because each OED’s criteria require a particular medical image processing task for detection. Therefore, we proposed a novel multi-task learning network to combine semantic segmentation and classification to detect OED using architectural and cytological characteristics. Our proposal is the first study that jointly trained semantic segmentation and classification on a single network for automated OED detection. We developed four new frameworks called VGG16-UNet, InceptionV3-UNet, DyspVGG16, and Dysp-InceptionV3. The VGG16-UNet and InceptionV3-UNet were designed based on classic U-Net with the ImageNet pre-trained VGG16 and InceptionV3 encoder and a traditional classifier model. We built Dysp-VGG16 and Dysp-InceptionV3 using our novel modified U-Net and novel classifier network. Our modified U-Net involved dilated convolution, channel attention, spatial attention, and residual blocks for performance enhancement. The proposed models’ effectiveness and robustness were verified by running three experiments and utilizing quantitative metrics and visualization results for comparison. Consequently, our novel modified U-Net and classifier network show superior performance on classification and segmentation tasks. Our novel classifier enhanced the quantitative metrics and reduced the traditional classifier’s false positives and negative rates. Modified U-Net improved the semantic segmentation performance by 5% of the Jaccard index and provided accurate predicted masks.