Cellular Level Based Deep Learning Framework for Early Detection of Dysplasia in Oral Squamous Epithelium

摘要

Over the past few decades, the artificial intelligence is being employed in diverse fields like pattern classification, image processing, object identification, rec-ommender systems, speech recognition, etc. Machine learning has made it possible to develop intelligent systems through training that equip machines to handle different tasks, exactly on the analogy similar to humans. In medical field, machine learning algorithms are being used for prediction, early detection and prognosis of various diseases. These algorithms suffer a certain threshold due to their inability to handle large amount of data. Deep learning based techniques are emerging as efficient tools and can easily overcome the above difficulties in processing data related to medical imaging that includes mammographs, CT scans, MRIs and histopathology slide images. Deep learning has already achieved greater accuracy in early detection, diagnosis and prognosis of various diseases especially in cancer. Dysplasia is considered to be a pathway that leads to cancer. So, in order to diagnose oral cancer at its early stage, it is highly recommended to firstly detect dysplastic cells in the oral epithelial squamous layer. In our research work, we have proposed a deep learning based framework (convolutional neural network) to classify images of dysplastic cells from oral squamous epithelium layer. The proposed framework has classified the images of dysplastic cells into four different classes, namely normal cells, mild dysplastic cells, moderate dysplastic cells and severe dysplastic cells. The dataset undertaken for analysis consists of 2557 images of epithelial squamous cells of the oral cavity taken from 52 patients. Results show that on training the proposed framework gave an accuracy of 94.6% whereas, in testing it gave an accuracy of 90.22%. The results produced by our framework has also been tested and validated by comparing the manual results recorded by the medical experts working in this area.