Observational Variability in Neural Networks for the Classification of Oral Epithelial Dysplasia

The diagnosis of Oral Epithelial Dysplasia (OED), presents a high interobserver variability due to subjectivity in the evaluation criteria. In this work, we propose an automatic histological image classification strategy based on the multiple instance learning (MIL) approach, using VGG-16 convolutional neural networks for feature extraction. Four models were trained: Two for classifying the degree of OED (mild, moderate, and severe) and two for the detection of six relevant histopathological criteria. To optimize the training process, we implemented the Black Hole metaheuristic to find the learning rate that maximizes the performance of the models. Evaluation of performance and interobserver variability was performed using Cohen's Kappa coefficient. The results suggest that the use of MIL, together with metaheuristic optimization strategies, can consistently reproduce expert diagnostic perception.