CAMERA‐Net With Focal Pixel Loss for Accelerated MRI

The slow acquisition time of MRI poses a significant barrier to its widespread clinical adoption. To address this limitation, we introduce CAMERA-Net, a novel multi-stage cascaded wavelet neural network for reconstruction of accelerated MRI. CAMERA-Net incorporates a wavelet transform-based regularization network that effectively leverages wavelet transforms and convolutional neural networks to preserve critical image information. The cascaded structure of CAMERA-Net offers flexibility in adjusting its depth without increasing the model size. Additionally, we propose a focal pixel loss strategy to dynamically adjust pixel-wise loss weights, enabling the network to prioritize challenging pixels and enhance reconstruction accuracy. Our findings underscore the superior performance of the wavelet transform-based regularization network compared to U-net, the advantages of the cascaded structure, and the effectiveness of the focal pixel loss strategy in improving reconstruction quality. Experimental results on the fastMRI dataset demonstrate that CAMERA-Net outperforms existing state-of-the-art methods. CAMERA-Net provides a promising framework for accelerated MRI, offering both accuracy and flexibility.