Abstract
Background: The Yamamoto-Kohama criteria (YK), which are used to classify the morphology of the infiltrating protrusions of an oral squamous cell carcinoma (OSCC), are clinically useful for determining the mode of tumor invasion, especially in Japan. However, evaluations of the mode of OSCC invasion are based on subjective visual observations, and this approach has created considerable inter-evaluator and inter-facility differences. In this retrospective study, we aimed to develop an automatic method of determining the mode of invasion of OSCC based on the processing of digital medical images of the invasion front. Methods: Using 101 digitized photographic images of anonymized stained specimen slides from consecutive patients with OSCC at Kanazawa University, we created a classifier that allowed clinicians to introduce feature values and subjected the cases to machine learning using a random forest approach. We then compared the YK grades (1, 2, 3, 4C, 4D) determined by a human oral and maxillofacial surgeon with those determined using the machine learning approach. Results: The input of multiple test images into the newly created classifier yielded an overall F-measure value of 87%, (Grade 1: 93%, Grade 2: 67%, Grade 3: 89%, Grade 4C: 83%, Grade 4D: 94%). These results suggest that the output of the classifier was very similar to the judgments of the clinician. Conclusions: We successfully developed an automatic machine-learning method for discriminating the mode of invasion of OSCC. Our results suggest that a medical diagnostic imaging system could feasibly be used to provide an accurate determination of the mode of OSCC invasion.