Classification of oral lesions using deep learning for early detection of oral cancer

Lead Research Organisation: Kingston University
Department Name: Fac of Science Engineering and Computing

Abstract

For the majority of cancers, early detection results in better survival. Oral cancer is one of the few cancers that is visible and many of these cancers are preceded by a potentially malignant lesion where medical intervention can prevent the development of cancer. Taken together, oral cancer presents an opportunity for early detection. However, identifying which oral lesion has a propensity to become oral cancer is not straightforward without specialised training and this problem is confounded by the lack of specialists who are trained in this expertise particularly in low- and middle-income countries, where the majority of oral cancers are diagnosed. One innovative approach to overcome this is to develop an artificial intelligence algorithm to classify oral lesions into those that are benign and those that are potentially malignant or are occult cancer so that patients can be triaged accordingly to receive appropriate clinical management. In this project, we propose to work within a multi-disciplinary, international team to collate a library of images from existing and prospective collections that will facilitate the development of an artificial intelligence algorithm that will be tested and validated. The outcome of this project will pave the way for further rigorous testing, development of an App incorporating this automated tool and clinical validation for the early detection of oral cancer. The development of an automated tool for the classification of oral lesions will facilitate the identification of patients most at risk to develop oral cancer so that these individuals can be managed appropriately. This project is particularly impactful in the low- and middle-income countries as the majority of the global burden of oral cancer is found in these countries.

Technical Summary

About 300,000 individuals are diagnosed with oral cancer every year and the majority of these patients live in low- and middle-income countries. Up to 80% of oral cancers are preceded by early lesions called oral potentially malignant lesions (OPMD) and this presents an opportunity to identify and manage these lesions to reduce the chances of malignant transformation. However, due to the numerous types of lesions that present in the oral cavity, the identification of OPMD is not straightforward. Further, this is confounded by the limited number of trained oral medicine specialists in countries where the prevalence is at its highest.

One way to overcome the lack of access to expertise for the clinical diagnosis of these lesions is to use deep learning approaches to develop an algorithm that could automatically classify images into different categories that require distinct clinical management. Such an approach has been successfully used for the classification of skin diseases. In this project, we will work within a multi-disciplinary team to use deep learning for the classification of oral lesions. In particular, we will employ two different deep learning architectures - VGG-16 and ResNet, pre-trained using ImageNet, and apply transfer learning to classify oral lesions.

The outcome of this project will pave the way for further rigorous testing, development of an App incorporating this automated tool and clinical validation for the early detection of oral cancer. The development of an automated tool for the classification of oral lesions would facilitate the identification of patients most at risk of developing oral cancer so that these individuals can be managed appropriately.

Planned Impact

This project has benefits to different stakeholders:
Clinicians and Ministries of Health: An automated tool that could help with patient triaging to streamline healthcare would be very useful for primary healthcare practitioners to manage their patients. This would also ensure that specialists are seeing patients that are most at risk, streamlining resources that are often limited in low-and middle-income countries. As oral cancer is very often treated in the public hospitals, such an automated classification tool would have the farthest reach through the Ministry of Health. In Malaysia, we are already partnering with the Ministry of Health to conduct clinical trials on MeMoSA and the same network will be involved in this project. This strong and long-lasting partnership will facilitate the testing and use of the automated tool in the longer term and together with other research, including health economics, data from this project may play a role in shaping healthcare policies in Malaysia and other countries in the region.

Communities in low- and middle-income countries and globally: An automated tool for early detection of oral cancer would be most beneficial to low- and middle-income countries where there are not enough specialists to diagnose patients. We have already demonstrated that a mobile phone App can be used to facilitate early detection of potentially malignant disorders in the oral cavity, and internet connectivity in low- and middle-income countries permits the use of telemedicine. Taken together, combining these (automated classification and telemedicine) would be an innovative and cost-effective approach to address the challenges faced by low- and middle-income countries.

Commercial partners/Social enterprise: To facilitate the wide use of a successfully developed tool for early detection of oral cancer would require a commercial partner. Therefore, in the longer-term, discussions with commercial partners would have to take place or team members may decide to set up a social enterprise so that funds can continue to finance further projects. Intellectual property will be co-owned by the investigators of this project. Currently such a tool is not available in the market and therefore there is a large potential for commercialization.

Academic/Research communities: This project will be one of the first to provide evidence that artificial intelligence approaches can be used for the classification of oral lesions and will stimulate the field for further development of deep neural networks for other diseases involving medical imaging, and could inspire the development of multi-disciplinary teams to address healthcare issues. Further, this research would also expand other areas of research including health economics to determine the cost-savings as a result of implementing an automated classification tool for the early detection of oral cancer. Data and results from this study will be disseminated at scientific conferences and through publication.

Publications

10 25 50