Subband Fusion Based W-GAN for Biomedical Image Super-Resolution

Finding tiny abnormality, textual deviation and other details are crucial for diagnosis in medical images. Imageprocessing resulting into super-resolution high quality medical imaging helps clinicians for accurate analysis. Medical images are acquired using different modalities and developing a common method for biomedical images of different modalities is immensely challenging. Popular biomedical imaging modalities, are CT, PETCT, MRI, and ultrasound etc. Super-resolution achieved through iinterpolation-based techniques are often reported for its quality to computational complexity trade-off. Learning based upsampling using transposed convolution and sub-pixel layers are anticipated overcome the shortcomings of interpolation method. SRGAN and ESRGAN models have recently become popular generative models for their performance in superresolution. This paper proposes a novel Wavelet GAN (W-GAN). Proposed method hasbeen tested on medical images from different modalities. Results are compared for the conventional interpolation method, ESRGAN, and the proposed MS-SRGAN. Though Deep Super resolution itself is a multiscale approach, the innovation of the proposed work is decomposing the input low resolution image into subbands by using discrete wavelet transform. Therefore, deep representation along with separated complementary textual and edge features has been proven efficient. Training of GAN models gets streamlined and efficient for images of different resolution and having similar high frequency information. Reconstructed SR image obtained from inverse discrete wavelet transform has better clarity than ground truth image. The parameters like Peak signal to noise ratio (PSNR) and structural index similarity (SSIM) are used to quantify the improvement in super resolute images. In our proposed method we observed that the quantitative result is better in terms of PSNR and SSIM as compared to state of art methods.