The study on the impact of activation function and pooling strategies on U-Net-based medical image fusion of traumatic brain images

Image fusion is the technique in which two or more images are merged to enrich the information content of the image. In this study fusion of CT and MRI images is carried out using U-net architecture-based CNN model. In CNN, activation function and pooling strategies play a main role in determining the quality enrichment of fused images. Activation functions (AF) provide non-linear learning, helping build robust training models with a wide range of feature extraction improving the model performance. Pooling is a technique used in CNN to down-sample the image and to improve the spatial dimension by retaining the information in the image. In this study, we aim to identify the combinations of activation function and pooling strategy to improve the neural network performance, we inspect the effectiveness of the ReLU and Sigmoid functions with a combination of max pooling and average pooling techniques. These combinations are evaluated using a series of metrics: Peak Signal Noise Ratio (PSNR), Structure Similarity Index (SSIM), Mean Squared Error (MSE), Entropy, Fusion Factor, Mutual Information (MI), and Edge Preservation with the best values of 70.0073, 0.8547, 15.61, 1.2205, 0.8546, 0.2559, 0.3388 respectively. The results imply that the ReLU activation function, when united with max pooling, consistently improves the performance of the U-Net model in most measures. This combination provides better performance extraction and preserves important details in the processed image.