Efficient Deep Learning for Melanoma Diagnosis on CPU Hardware

Convolutional Neural Networks (CNNs) are a category of Neural Networks that have grown increasingly prominent in Computer Vision applications such as Image Classification, Image Segmentation, and Object Recognition. Practical uses of CNNs are seen in many academic fields, this study focuses on exploring their applications in Medical Imaging, particularly classifying skin lesions into benign or malignant cases using deep learning techniques to improve diagnostic accuracy. Along with the neural network architecture containing four convolutional layers, the model implements ReLU activation, max-pooling, and batch normalization to improve the accuracy and efficiency of the model. Due to the binary nature of the model's classification, an adaptive average pooling layer, dropout regularization, and a sigmoid activation function were also employed. The training and validation process was done on a dataset consisting of 10600 images of Melanoma, publicly available under the CC0 Domain License on Kaggle. The loss function was calculated using Binary Cross Entropy Loss. This model incorporated data augmentation techniques such as random rotations, flipping, and affine transformations for improved generalization. After 30 epochs, the network achieved 90.69% training accuracy and 91.20% test accuracy, with notable computational efficiency processing approximately 106.7 images per second on CPU hardware. The total runtime of both train and validation was approximately 45 minutes, highlighting its efficiency. The test accuracy was visualized via a confusion matrix, which validated the reliability of its predictions. These results underline the potential of this CNN as an automated clinical decision-support tool for dermatologists in aiding the early and accurate diagnosis of melanoma, proving how powerful deep learning can be in solving complex medical imaging challenges.