Pendekatan Multi-Input dalam Deteksi Kanker Kulit: Implementasi EfficientNetV2-B2 dan LightGBM

Skin cancer is one of the types of cancer with a high prevalence rate, so early detection is very important to increase the chances of recovery. This study aims to develop a skin cancer detection model that combines image data and tabular data using EfficientNetV2-B2 for image feature extraction and LightGBM for tabular data prediction estimation. The ISIC 2024 dataset used consists of 401,059 images of skin lesions with tabular features, including age, gender, location, diameter, and shape of the lesions. Tabular data is processed with normalization and encoding to avoid bias. Image data is also processed with augmentation techniques from kerascv. This multi-input model combines image and tabular features using concatenation techniques, with a dense layer as the final output. Our findings show that the model's accuracy and AUC value reached 96% and 98%, with success in handling class imbalance using undersampling and oversampling techniques. This study shows that the combination of images and tabular data increases the accuracy of skin cancer detection by 2%, compared to conventional CNN models, which only achieve an accuracy of around 94%. Moreover, this model offers better computational efficiency compared to conventional CNN models. The main contribution of this research is the use of multi-input that complements visual information with clinical data for more accurate and efficient skin cancer detection.