Image reconstruction is a key task in various applications related to medical scanning, geological prospecting, etc. Usually measured data for reconstruction is limited due to physical constraints in these practices. Consequently image reconstruction with limited data becomes an underdetermined inverse problem. Recently, using un-trained Deep Neural Networks (DNN) such as Deep Decoder and ConvDecoder has been proved as a plausible approach for image reconstruction. By leveraging the DNN structure itself as a prior, the un-trained neural networks have shown great advantages over the traditional reconstruction methods like total-variation (TV) in many aspects. Comparing with the stateof-art trained neural networks like U-net, the un-trained neural networks also alleviate the cost of robustness and scalability in out-of-distribution samples. In this work, we propose a new un-trained image reconstruction neural network called EffiDecoder. Our contribution is two-fold. First we introduce mobile inverted bottleneck (MBconv) at each layer with computational-friendly depthwise separable convolutions. Second, the feature channels are adaptively reduced according to the output. In accelerated MRI reconstruction tasks the experimental results demonstrate that the proposed method can achieve competitive performance with reduced computational cost. Therefore, with the notable improvements of EffiDecoder, we believe the un-trained neural networks have great potentials to be discovered in extensive reconstruction applications.
|