Skip to content

Latest commit

 

History

History
43 lines (37 loc) · 2.36 KB

README.md

File metadata and controls

43 lines (37 loc) · 2.36 KB

Artificial Intelligence in Medical Imaging

This project was done as part of a 4th year capstone project at the University of Toronto to reduce time and space demands of MRI scans. We designed a system using U-Nets and kMeans to optimally select a subset of k-space points from a MRI scan that yields a high quality image. Our system utilizes a custom masking technique and a modified version of Facebook’s fastMRI U-Net to quickly reconstruct a MRI image from a subset of k-space points. The system overall is able to achieve Structural Similarity Index Metric scores exceeding 75% between the ground truth and the reconstructed MRI image.

Usage

In order to use or build upon the existing code, the custom filtering technique is provided in KSpaceReductionModel.ipynb, you will need a set of MRI scans and the Facebook's fastMRI, both of which are available here.

Context

Current Medical Resonance Imaging (MRI) reconstruction procedure is inefficient; k-space data acquisition and transformation is time consuming

Goal

Determine if machine learning algorithms can be applied to identify a subset of k-space points from an MR scan, that when reconstructed produces an MR image of a comparable level of detail to that of the ground truth image.

Design Flow

Alt text

  1. A raw k-space dataset is first fed into the k-space reduction model to obtain a sub-sampled k-space dataset via unsupervised learning models like K-Means clustering.
  2. A U-Net model is pre-trained on 300 k-space subsets (i.e. 9000 images) to accurately enhance the input subsampled k-space dataset
  3. During the prediction phase, we send the subsampled k-space dataset from part 1 through the U-Net to obtain an enhanced MR image
  4. SSIM is then computed for ground truth and reconstructed image for quantifying similarity
  5. For our experiment, we calculated the mean SSIM of the 1050 images that were obtained from the U-Net. Our final SSIM score was 0.80285.

Results and Conclusion

  • Achieved final SSIM of 80.285%, beating our target of 75%
  • Confirmed hypothesis of patterns in k-space, obtainable via AI algorithms
  • Fulfilled all project requirements, while maintaining flexibility through tunable parameters

Alt text Alt text