Sub-RENet: a wavelet-based network for super resolution of diagnostic ultrasound

In past few years artificial intelligence in super resolution (SR) has taken a big leap. However, there is still scope of improvement for SR of diagnostic ultrasound (US) images. So, we have proposed a Sub-band Resolution Enhancement Network (Sub-RENet) for SR of the diagnostic US images. The efficiency of wavelets in multiresolution analysis, inspired us to use wavelet sub-bands for training. Sub-RENet has two modules B-net and A-net. B-net takes the resolution to next scale while A-net takes it to higher scales. Both networks use residual and skip connections and unique set of loss function. We proposed a weighted multi-network multi-scale perceptual loss to preserve image detail. The self-supervised A-net uses distribution loss to avoid artifacts. The B-net had average mean squared error and structural similarity of 628 and 0.79 for breast US image dataset and 459, and 0.82 for liver US image dataset, respectively. For output of ×8 input resolution, the values of blind image spatial quality evaluator and just noticeable blur were 43.3 and 0.875, for breast US image dataset and 43.7, and 0.873 for liver US image dataset, respectively, which is far better than recent SR networks. The consistency in values of evaluation metric for two different diagnostic US images shows the applicability of Sub-RENet in any diagnostic US images.