Figure S3.

The Residual Neural Network model, with a random cropping and voting strategy, achieves an F-score of 0.87.

(A) The training loss curves of the Vision Transformer model (left) show the convergence of the training and validation loss, highlighting the ideal model condition. The X-axis, labeled “Epochs,” represents the number of training attempts, while the Y-axis, labeled “Training Loss,” represents the magnitude of error after each run. The F1 score curves of the Vision Transformer model (right) show the validation score in each epoch. (B) Schematic flowchart of the inputs, ResNet-50 model, and growth size classification. Two different inputs were fed to the model. Input #1 was the original palmskin images; Input #2 had random cropping, a data augmentation technique, applied to the palmskin images to generate a larger set of imaging data for the model. (C, D) Confusion matrix and F-scores for the “No voting” and “Voting” conditions. Scores above 0.7 are shown in black for ease of visualization.