Beyond Dysplasia: Uncovering Structure in Oral Potentially Malignant Diseases with Unsupervised Contrastive Learning

Automated cancer diagnosis research often focuses on a binary task - recognize dysplasia and cancer from other lesions. However, other clinical conditions have estimated malignant transformation rates. Grouping these oral potentially malignant diseases with benign conditions may not be ideal. While automated cancer diagnosis potential has been shown in multi-spectral autofluorescence images, the existence of a more detailed structure has not been investigated. Training multi-class models is often avoided on small data sets due to the low number of samples per class, but standard clustering algorithms like k-means lack supervision to manage high inter-patient variability. We propose an unsupervised contrastive clustering approach, where a small neural network is trained to group normal images together, away from lesion images. We expect our method to reduce inter-patient variability and amplify the existing structure among different types of oral lesions. Our results indicate that a structure based on risk level for malignant transformation does exist, generating a cluster composed of over 75% high-risk lesions and out-performing the k-means baseline of 40%.Clinical relevance- Instead of focusing on the identification of dysplasia and squamous cell carcinoma from other lesions in a binary task, our investigational analysis shows that more detailed structure exists in multi-spectral autofluorescence images, with applications in the management of oral potentially malignant diseases.