Fang Liu

This talk will provide an overview of the latest deep learning techniques that have been applied to image reconstruction of musculoskeletal MRI. The topics focus on rapid imaging for both static MRI and quantitative MRI, such as T2 and T1rho mapping. The talk will conclude by highlighting potential challenges in deep learning-based reconstruction that warrant further investigation.

Cem Deniz

This educational lecture will provide an overview of the machine learning approaches applied in MR image processing and interpretation for musculoskeletal disorders.

Francesco Caliva¹, Adam Noworolski², Andrew Leynes^1,3, Claudia Iriondo^1,3, Sharmila Majumdar¹, Peder Larson¹, and Valentina Pedoia^1
1Department of Radiology and Biomedical Imaging, University of California, San Francisco, San Francisco, CA, United States, ²EECS, University of California, Berkeley, Berkeley, CA, United States, ³Department of Bioengineering, University of California, Berkeley, Berkeley, CA, United States

We propose a novel task based deep learning framework for simultaneous MRI reconstruction and segmentation. On a dataset of retrospectively undersampled knee-DESS volumes we demonstrate that irrespective of ultra-high acceleration factors (i.e. 48×) a multitask 3D encoder-decoder is capable of reconstructing with high fidelity the knee MRI, accurately segment cartilaginous and meniscal tissues and reliably provide cartilage thickness. Our multitask solution outperforms two other methods: a compressed sensing reconstruction step, followed by a deep learning-based tissue segmentation. The other method comprises a cascade of two convolutional neural networks that sequentially perform image reconstruction and segmentation.

Radhika Tibrewala¹, Eugene Ozhinsky¹, Rutwik Shah¹, Io Flament¹, Kay Crossley², Ramya Srinivasan¹, Thomas M Link¹, Valentina Pedoia¹, and Sharmila Majumdar^1
1Radiology and Biomedical Imaging, University of California San Francisco, San Francisco, CA, United States, ²La Trobe Sport and Exercise Medicine Research Centre, Melbourne, Australia

MRI based hip degeneration grading is difficult, time-intensive and prone to inter-reader variability, aggravated by the lack of a standard hip grading scale. Recent research using deep learning based clinical classification tasks has shown efficiency in knee degenerative changes. In this study, we aim to develop a deep learning based hip degenerative changes classification model (for cartilage lesions, bone marrow edemas and cysts) and evaluate its performance. In addition to that, we develop an AI-assist tool based on model predictions to test on two radiologists to see if the inter-reader agreement increases by using the AI-assist.

Shui Liu¹, Fei Gao², Xiaodong Zhang¹, Jue Zhang^2,3, Xiaoying Wang^1,3, and Jing Fang^2,3
1Department of Radiology, Peking University First Hospital, Beijing, China, ²College of Engineering, Peking University, Beijing, China, ³Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, China

To enrich the representation capability of the CNN model and achieve more accurate lumbar disc degeneration grading, inspired by student T-test in statistics, we propose a T-test regularization strategy focusing on pushing away different categories from each other in feature space.

Hongyu Li¹, Mingrui Yang², Jeehun Kim², Ruiying Liu¹, Chaoyi Zhang¹, Peizhou Huang¹, Sunil Kumar Gaire¹, Dong Liang³, Xiaojuan Li², and Leslie Ying^1
1Department of Biomedical Engineering, Department of Electrical Engineering, The State University of New York at Buffalo, Buffalo, NY, United States, ²Program of Advanced Musculoskeletal Imaging (PAMI), Cleveland Clinic, Cleveland, OH, United States, ³Paul C. Lauterbur Research Center for Biomedical Imaging, Medical AI research center, SIAT, CAS, Shenzhen, China

This abstract presents a deep learning method to generate MR parameter maps from very few subsampled echo images. The method uses deep convolutional neural networks to learn the nonlinear relationship between the subsampled T1rho/T2-weighted images and the T1rho/T2 maps, bypassing the conventional exponential decay models. Experimental results show that the proposed method is able to generate T1rho/T2 maps from only 2 subsampled echo images with quantitative values comparable to those of the T1rho/T2 maps generated from fully-sampled 8 echo images using the conventional exponential decay curve fitting.