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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oral Pathology

Background:

• Automated classification of oral lesions using Convolutional Neural Networks (CNNs) is an emerging field.
• Understanding CNN error patterns is crucial for improving diagnostic accuracy in oral medicine.

Purpose of the Study:

• To analyze errors from a CNN model classifying oral lesions based on clinical characteristics.
• To identify systemic error patterns within the intermediate layers of the CNN.

Main Methods:

• Retrospective analysis of a CNN model's classification errors on oral lesion images.
• Cross-sectional analysis nested within a previous trial.
• Examination of 116 misclassified outputs (7.6% of total images).

Main Results:

• Discrepancies were associated with image sharpness, resolution, and focus.
• Human errors and the impact of data augmentation significantly influenced classification accuracy.
• Qualitative analysis confirmed the effect of image quality and data augmentation.

Conclusions:

• Image quality and data augmentation are critical factors affecting CNN performance in oral lesion classification.
• Understanding CNN decision-making factors enhances confidence in their diagnostic potential.
• Further refinement of CNN models can improve automated oral lesion diagnosis.