Abrams Presentation

Transcript

1. Development of Deep Learning System for Quantitative MRI of the Knee Jed Tan Mentor: Dr Lalush
2. Background Information • After ACL reconstruction, the changes in cartilage are not completely known. • Research projects targeting the articular cartilage utilize manual segmentation for image registration. Manual Segmentation • A project worked on by Dr. Lalush studied articular cartilage manually segmented knee images across 20 subjects [1] • Other projects have created programs capable to automatically segmenting bones [2] Segmentation: image is broken down into subgroups, called segments 2D knee MRI image Segmented 2D knee MRI image
3. Objective Develop a deep learning neural network that automatically segment bones. Our long term goal is to understand how cartilage changes as a function of location under different conditions. Our Approach: 1. Prepare training data for the neural network a. Manually segment knee MRI images b. Crop 3D images to same number of slices c. Flip left knees to match right side knees 2. Design the network a. Unet architecture 3. Train the network a. Use of high powered graphics cards b. Compare loss Manual Segmentations Segmented 3D Image Segmented 2D Image
4. Unet Architecture
5. Results Figure 1. Loss of Validation Set Figure 2. Loss on Training Set
6. Discussion • The network has low loss values, but it is not perfect Certain slices may have parts that are segmented incorrectly or have noise • After using the network the segmented 3D images should still be manually checked over and touched up • This technology is unique to our MRI imaging sequences and will be used in several projects ◦ Network was deployed to UNC Department of Exercise and Sports Science ◦ Used to segment bones in Andrew Marshall’s knee project
7. Conclusion • A deep learning neural network has been successfully developed that can automatically segment bones and has been put to use • Our next project is to develop a network to automatically segment the cartilage • From here we can start looking into strain maps of the changes in cartilage mapped on the femur and tibia Thank you to the Abrams Scholars Program for providing me this opportunity and Dr. Lalush for mentoring me through this project. Data for this project was funded by grants from the Arthritis Foundation and UNC Thurston Arthritis Research Center.
8. References [1] Boling MC, Dupell M, Pfeiffer SJ, Wallace K, Lalush D, Spang JT, Nissman D, Pietrosimone B. In Vivo Compositional Changes in the Articular Cartilage of the Patellofemoral Joint Following Anterior Cruciate Ligament Reconstruction. Arthritis Care Res (Hoboken). 2021 Jan 18:10.1002/acr.24561. doi: 10.1002/acr.24561. Epub ahead of print. PMID: 33460530; PMCID: PMC8286261. [2] Shen, W., Xu, W., Zhang, H., Sun, Z., Ma, J., Ma, X., Zhou, S., Guo, S., & Wang, Y. (n.d.). Automatic segmentation of the femur and tibia bones from X-ray images based on pure dilated residual U-Net. https://doi.org/10.3934/ipi.2020057