AN ENHANCED EFFICIENTNETB0 WITH CBAM ATTENTION FOR ROBUST SKIN CANCER DETECTION

Abstract

Skin cancer is one of the most common and deadly types of cancer in the world, so it's important to get the right diagnosis quickly. In this study we propose a strong deep learning framework based on EfficientNetB0 and the Convolutional Block Attention Module (CBAM). This will help the model better focus on important areas in dermoscopic images. The model is trained and tested on a dataset of skin cancers containing benign and malignant classes. The size of the dataset for training is 2637 images, and 660 images for the test set. The dataset was obtained from Kaggle. The EfficientNetB0 was selected as the backbone architecture because of the small number of parameters. The batch normalization and dropout layers were used for the model to generalize better and stop it from overfitting. The proposed study CBAM block for the attention mechanism.  Grad-CAM visualizations show which parts of an image affect the model's predictions, which makes the decision-making process more open and trustworthy. The model achieved 90% accuracy. The proposed method shows great promise for use in automated skin cancer screening systems and can help dermatologists.