Dense Net-Based Acute Lymphoblastic Leukemia Classification and Interpretation through Gradient-Weighted Class Activation Mapping

Deep learning models have shown remarkable accuracy in Acute Lymphoblastic Leukemia classification tasks, but their decision-making process remains challenging to interpret. Gradient-weighted Class Activation Mapping (Grad-CAM)is employed to address this opacity. By visualizing the regions of input images that contribute most to predictions, this project provides insights into the decision-making rationale of the models. The approach of this project involves generating augmented data using various techniques such as flipping, rotating, and zooming to remove data imbalance in a dataset. After that, split the data into training and testing data, fine-tune the DenseNet model on the trained data, and evaluate the performance of both models using testing data based on Area Under Curve, recall, accuracy, precision, and f1-score, then Integrate the model into a grad- CAM to visualize the key image region. The model achieves an accuracy of 94.881 %. Comparative performance metrics, including Area Under Curve, recall, accuracy, precision, and fl-score, highlight the superiority of the DenseNet model over MobileNet and EfficientNet B3 for ALL classification. This suggests that DenseN et is particularly well-suited for handling the complexities of the CNMC Leukemia dataset. By visualizing important regions in the images, Grad-CAM provides users with a comprehensive understanding of the decision-making process within the DenseNet model.