DEEP LEARNING AUGMENTED ASSESSMENT OF SKIN PHOTODAMAGE INFORMED BY MULTIPLE DERMATOLOGISTS

Non-melanoma skin cancers are primarily caused by ultraviolet radiation and affects a large population of the United States. The only available tool to assess skin photodamage is the McKenzie scale. However, the subjective and qualitative nature of this method leads to variability and inconsistency among dermatologists. We propose applying a deep learning approach to address this issue. 55 patients were assessed by 15 board-certified dermatologists rating the degree of skin photodamage using the McKenzie scale. Using a pretrained convolutional neural network, we train and test a model on labeled forearm images classified based on the severity of photodamage. We employ image preprocessing and data augmentation to the dataset as well as configure parameters and hyperparameters of the network architecture to obtain the optimal model to predict the degree of photodamage on the skin. Cross validation is performed to ensure the practical effectiveness of the model. Finally, performance of the neural network model is compared to that of the dermatologist ratings to determine feasible application of this model. We envision this as augmented technology for objective and reliable assessment of skin photodamage for dermatologists.

skin_damage.zip contains the original collected images from the clinical study from Wright State University. Segmentation_v3.m segments the images in N number of squares based on user input, with user-dictated adjustments for rotation, cropping, and mask tolerance. Currently, N = 4 and can be adjusted accordingly. Executing move_files.m segregates the segmented images into their respective categories within the appropriate testing and training data for cross-validation. rename_files.m then renames those images to the proper naming convention.

Crossval.zip is an example of the output of these steps based on . pdnn.m executes the neural network, and accuracy_check_v3.m results in a confusion matrix of the classified data.