SKIN CANCER DETECTION USING THE CONVOLUTIONAL NEURAL NETWORK METHOD WITH MOBILENETV3 ARCHITECTURE ON DERMATOSCOPIC IMAGES

Abstract

Skin cancer is one of the most common types of cancer with a continuously increasing global incidence rate, making early detection essential to improve treatment success and reduce mortality. This study aimed to develop a mobile-based skin cancer detection application using the Convolutional Neural Network (CNN) method with the MobileNetV3 architecture implemented in a Flutter application. The dataset used consisted of 1,500 dermatoscopic images divided into three classes: Basal Cell Carcinoma (BCC), Squamous Cell Carcinoma (SCC), and healthy skin (Normal). The research stages included data preprocessing through resizing, normalization, and image augmentation, model training using MobileNetV3Small with transfer learning and fine-tuning techniques, model conversion into TensorFlow Lite (TFLite) format, and implementation into an Android-based mobile application. Model evaluation was performed using a confusion matrix with accuracy, precision, recall, and F1-score metrics. The results showed that the model achieved an accuracy of 93.75% with average precision, recall, and F1-score values of 0.94. In single-image prediction testing, the model achieved confidence scores of 98.44% for the BCC class, 98.14% for the SCC class, and 99.91% for the Normal class. The implementation of Test Time Augmentation (TTA), Dropout, and L2 regularization proved effective in improving model stability and performance. In addition to its high accuracy, the application was able to run in real time on Android devices with fast inference time and efficient memory usage. This study demonstrates that the combination of MobileNetV3Small and Flutter can provide a lightweight, accurate, and accessible solution for early skin cancer detection.