AI Grading of Lateral Canthal Lines: Novel Models for Unseen Synthetic Image Generation and Data Augmentation.

Purpose: Large and balanced datasets are required to train artificial intelligence (AI) algorithms but are often difficult to acquire using clinical dermatologic photographs. We aimed to develop a new diffusion-based generative AI algorithm that generates patient photographs with modifiable details, thereby creating large and balanced datasets for neural network training. The newly developed model was tested using lateral canthal lines, a common patient's reason for seeking treatment. Patients and Methods: Five hundred and sixty-six photographs of the lateral oblique face of graded by certified dermatologists according to the severity of lateral canthal lines. We developed a zero-shot and few-shot image generation model that adds structured compositional labels as control variables to the diffusion model. This allows us to create synthetic images from original photos while adjusting the severity of lateral canthal lines with only a few examples. The generated images were used to train a convolutional neural network (CNN) with ResNet-34 backbone for classifying the grade of lateral canthal lines. Results: We successfully generated 10,500 patient images similar to the original photographs with different grades of lateral canthal lines. The accuracy (82% vs 91%) and the area under the receiver operating characteristic curve (0.935 vs 0.981) of the classification CNN remarkably improved after training with the new dataset containing generated images. Conclusion: The compositional zero-shot and few-shot generation model is able to generate images similar to original clinical photographs, and the features of the images can be modified to match the needs of the specific task, allowing researchers to create a larger and more balanced dataset to improve neural network training outcomes. This is especially important in dermatology, where large-scale clinical photographs are difficult to acquire for machine learning. The results of this study are limited by low patient diversity and a lack of external validation.