Convolutional neural network misclassification analysis in oral lesions: an error evaluation criterion by image characteristics.

OBJECTIVE: This retrospective study analyzed the errors generated by a convolutional neural network (CNN) when performing automated classification of oral lesions according to their clinical characteristics, seeking to identify patterns in systemic errors in the intermediate layers of the CNN. STUDY DESIGN: A cross-sectional analysis nested in a previous trial in which automated classification by a CNN model of elementary lesions from clinical images of oral lesions was performed. The resulting CNN classification errors formed the dataset for this study. A total of 116 real outputs were identified that diverged from the estimated outputs, representing 7.6% of the total images analyzed by the CNN. RESULTS: The discrepancies between the real and estimated outputs were associated with problems relating to image sharpness, resolution, and focus; human errors; and the impact of data augmentation. CONCLUSIONS: From qualitative analysis of errors in the process of automated classification of clinical images, it was possible to confirm the impact of image quality, as well as identify the strong impact of the data augmentation process. Knowledge of the factors that models evaluate to make decisions can increase confidence in the high classification potential of CNNs.

Use of Artificial Intelligence in the Classification of Elementary Oral Lesions from Clinical Images.

OBJECTIVES: Artificial intelligence has generated a significant impact in the health field. The aim of this study was to perform the training and validation of a convolutional neural network (CNN)-based model to automatically classify six clinical representation categories of oral lesion images. METHOD: The CNN model was developed with the objective of automatically classifying the images into six categories of elementary lesions: (1) papule/nodule; (2) macule/spot; (3) vesicle/bullous; (4) erosion; (5) ulcer and (6) plaque. We selected four architectures and using our dataset we decided to test the following architectures: ResNet-50, VGG16, InceptionV3 and Xception. We used the confusion matrix as the main metric for the CNN evaluation and discussion. RESULTS: A total of 5069 images of oral mucosa lesions were used. The oral elementary lesions classification reached the best result using an architecture based on InceptionV3. After hyperparameter optimization, we reached more than 71% correct predictions in all six lesion classes. The classification achieved an average accuracy of 95.09% in our dataset. CONCLUSIONS: We reported the development of an artificial intelligence model for the automated classification of elementary lesions from oral clinical images, achieving satisfactory performance. Future directions include the study of including trained layers to establish patterns of characteristics that determine benign, potentially malignant and malignant lesions.