Deep Convolutional Encoder-Decoder algorithm for MRI brain reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Compressed Sensing Magnetic Resonance Imaging (CS-MRI) offers faster acquisition and improved image quality by reducing artifacts.
• Deep Learning (DL) shows promise in reconstructing high-fidelity MRI scans, enhancing parallel imaging techniques.
• Existing DL methods often require extensive training data, posing a challenge for clinical implementation.

Purpose of the Study:

• To develop and evaluate a Deep Convolutional Encoder-Decoder architecture for efficient and accurate CS-MRI reconstruction.
• To bridge the gap between traditional methods and DL approaches by utilizing a novel autoencoder architecture.
• To assess the model's performance on both normal and pathological MRI datasets, including noisy scans.

Main Methods:

• Proposed a Deep Convolutional Encoder-Decoder network, an autoencoder architecture with three convolutional blocks in both encoder and decoder.
• Evaluated the model using the Hammersmith (normal scans) and MICCAI 2018 (pathological MRI) datasets.
• Extended the model to handle noisy pathological MRI scans and compared its performance against state-of-the-art algorithms using NMSE, PSNR, and SSIM metrics.

Main Results:

• The proposed architecture demonstrated superior reconstruction quality, evidenced by higher Peak-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) values, and lower Normalized Mean Square Error (NMSE).
• The model effectively preserved textural image details in the reconstructed MRI scans.
• Achieved a running time of approximately 0.8 seconds, indicating suitability for real-time processing.

Conclusions:

• The developed Deep Convolutional Encoder-Decoder architecture provides effective CS-MRI reconstruction, outperforming existing methods in image quality and detail preservation.
• The model's efficiency and accuracy make it a viable tool for clinical settings, potentially improving patient experience and reducing costs.
• The architecture's ability to handle noisy scans further enhances its clinical applicability in diverse scenarios.