Superficial Wound Classification Using CNN Image Classifier for Initial Wound Care

Proper diagnosis of the types of wounds is the first step in effective wound management. Healthcare quality can be improved with the use of artificial intelligence technology. In this project, convolutional neural network (CNN) models are employed for superficial wound image classification. The goal of this project is to implement carefully selected CNN models with fine-tuned parameters (batch sizes, augmentation parameters and epoch size) to categorize superficial wounds of abrasion, bruise and superficial skin burns. The library and functions from TensorFlow, Keras, MatplotLib and OpenCv will be used alongside the Python programming language to create this project in a Google Colaboratory notebook, also widely known as Colab. This work evaluates the CNN models of VGG-16, InceptionV3 and ResNet-50 to find out which best classifies the dataset of non-fatal superficial wounds. These models have significant differences especially in their number of layers and concepts of capturing the features of images input. By utilizing these models, the accuracy rate of more than 40% is observed after models being trained and validated on 90% of the dataset (80% for training and 20 for validation), then tested and prediction with the remaining 10% of the dataset for classification report and confusion matrix generation. Out of 3 CNN models used in this study, it is observed that InceptionV3 model's highest accuracy is 44% while both ResNet50 and VGG-16 model achieved 45% accuracy against the superficial wound database (non-fatal wound).