Convolutional Neural Network-Based Clinical Predictors of Oral Dysplasia: Class Activation Map Analysis of Deep Learning Results

SIMPLE SUMMARY: Oral cancer/oral squamous cell carcinoma (OSCC) is among the top ten most common cancers globally; early and accurate diagnosis of oral cancer is critical. Despite improvement in surgical and oncological treatments, patient survival has not improved over the last four decades. Our purpose is to develop a deep learning method to classify images as “suspicious” and “normal” and to highlight the regions of the images most likely to be involved in decision-making by generating automated heat maps. Thus, by using convolutional neural network-based clinical predictors, oral dysplasia in an image can be classified accurately in an early stage. ABSTRACT: Oral cancer/oral squamous cell carcinoma is among the top ten most common cancers globally, with over 500,000 new cases and 350,000 associated deaths every year worldwide. There is a critical need for objective, novel technologies that facilitate early, accurate diagnosis. For this purpose, we have developed a method to classify images as “suspicious” and “normal” by performing transfer learning on Inception-ResNet-V2 and generated automated heat maps to highlight the region of the images most likely to be involved in decision making. We have tested the developed method’s feasibility on two independent datasets of clinical photographic images of 30 and 24 patients from the UK and Brazil, respectively. Both 10-fold cross-validation and leave-one-patient-out validation methods were performed to test the system, achieving accuracies of 73.6% (±19%) and 90.9% (±12%), F1-scores of 97.9% and 87.2%, and precision values of 95.4% and 99.3% at recall values of 100.0% and 81.1% on these two respective cohorts. This study presents several novel findings and approaches, namely the development and validation of our methods on two datasets collected in different countries showing that using patches instead of the whole lesion image leads to better performance and analyzing which regions of the images are predictive of the classes using class activation map analysis.