Brain tumor segmentation is performed using three-dimensional magnetic resonance images (MRIs) as a common practice. The accuracy and fast segmentation still remain a challenge. An algorithm is proposed using deep learning involving the UNet model with wavelet functions. In this algorithm, to enhance gradient flow and feature propagation, U-Net incorporates dense connections (residual block), which reduce the vanishing gradient issue and increase feature reuse. The encoder route blocks are extracted. Information is gradually abstracted, and the corresponding decoder path blocks reestablish the spatial resolution of the input images. The algorithm is validated on the data set MICCAI BratS2020 which includes a set of labelled brain tumor (2D MRI scans) dataset. A comparative study of UNet with Daubechies (db2,db4) and UNet with Haar tabulated for the Brats2020 dataset is included. The research design employed in this study combines residual blocks with UNet where the residual block is implemented using db2,db4 and Haar as activation function after batch normalization to improve training stability and accuracy. It was trained on 4308 2D MRI images, validated on 761 and tested on 895 MRI images.
Experimental results using db2 gave a validation accuracy of 0.99 with validation loss of 0.38 and a test accuracy of 0.98 with a test loss of 0.38 and CPU time taken was 9s 342 ms/steps.
Using db4 the validation accuracy was 0.99, with a validation loss of 0.4240 and test accuracy of 0.98 with test loss of 0.42 and CPU time noted was recorded as 7s 256ms/steps. For Haar wavelet a validation accuracy was 0.97, and loss wH 0.76 with test accuracy of 0.97 and loss of 0.75. Time taken was 6s 231 ms/step. It is noted that the deep learning strategy invoked along with db2 wavelet activation function shows remarkably improved accuracy and performance for segmentation as compared to its counterpart db4 and Haar wavelet family.