Identifying Oral Carcinoma from Histopathological Image using Unsupervised Nuclear Segmentation

Oral Cancer, a worldwide health concern, highlights the urgent need for accurate and swift detection and cure. Current diagnosing strategies primarily involve pathologists for analyzing tissue biopsy samples, a method that is time-consuming and heavily driven by pathologists' experience. To address these drawbacks, this study proposes a novel technique that incorporates machine vision for cancer detection, aiming to enhance diagnostic accuracy. Given the intricate nature of histopathological images, we adopt an unsupervised approach for cancer detection, in contrast to traditional deep learning or supervised approaches. The nucleus in a cancerous tissue biopsy image is identified as the region of interest (ROI), due to its key characteristics and form. We extract the ROI using K-means clustering augmented with a thresholding technique and apply a novel classification method for the final stage of cancer detection. Our proposed model achieved an accuracy of approximately 97.28%, with a closely following validation accuracy of roughly 96.34% making it more efficient and reliable at cancer detection. These results underscore the effectiveness of our two-stage process starting with image segmentation followed by CNN-based binary classification for accurately detecting cancer cells. They reveal improved speed and precision in identifying cancerous tissues, thus offering a promising pathway for enhancing the efficacy and efficiency of oral cancer detection.