Reducing Training Time in Skin Cancer Classification Using Convolutional Neural Network with Mixed Precision Implementation

In the field of skin cancer classification, machine learning and deep learning have been extensively utilized, particularly with convolutional neural network (CNN) architectures. However, there remains room for exploration to achieve optimal performance. This study investigates the use of the MobileNetV3Large architecture for transfer learning, chosen for its efficiency in low-power and memory-constrained applications. To further enhance performance, black-hat morphological transformation and oversampling techniques were applied to the ISIC 2020 dataset. Additionally, mixed precision training was implemented to reduce training time. The research aimed to compare the accuracy, precision, recall, F1-score, and training time of models trained with and without mixed precision. The findings revealed that while the model without mixed precision achieved superior performance with accuracy, precision, recall, and F1-score metrics reaching 98%, both models yielded an AUC-ROC of 1. Notably, mixed precision training significantly reduced training time by 1,646 seconds (27 minutes and 26 seconds), representing an 8.39% speed increase. These results suggest that mixed precision can meaningfully accelerate model training while maintaining competitive performance. The practical implications of this research include its potential to improve the efficiency of skin cancer classification models, making them more suitable for real-time clinical applications, particularly in resource-constrained environments.