Siyue LI¹, Shutian ZHAO¹, Yongcheng YAO¹, and Weitian CHEN^1
1AI in Radiology Laboratory, Department of Imaging and Interventioanl Radiology, The Chinese University of Hong Kong, Hongkong, Hong Kong

Accurate segmentation of the meniscus is valuable for clinical diagnosis and treatment of knee joint diseases. Due to expensive and time-consuming medical image data annotation, it is challenging to obtain sufficient labeled data for deep learning-based segmentation of meniscus. We investigated deep-learning based semi-supervised approaches with uncertainty estimation for meniscus segmentation using MR images. 

Dana Lin¹, Michael Schwier², Bernhard Geiger², Esther Raithel³, and Michael Recht^1
1NYU Grossman School of Medicine, New York, NY, United States, ²Siemens Medical Solutions USA, Princeton, NJ, United States, ³Siemens Healthcare GmbH, Erlangen, Germany

Rotator cuff tear size is a critical determinant of patient prognosis and surgical outcomes. Radiologists routinely make rotator cuff measurements as part of their MRI interpretation, which can be tedious and subject to variation among readers. This lends itself to a potential application for deep learning to increase efficiency and decrease variability in this task. In this study, we developed a DL model to generate measurements for full-thickness supraspinatus tendon tears.

Simon Sun¹, Ek Tsoon Tan¹, John A Carrino¹, Douglas Nelson Mintz¹, Meghan Sahr¹, Yoshimi Endo¹, Edward Yoon¹, Bin Lin¹, Robert M Lebel², Suryanarayan Kaushik², Yan Wen², Maggie Fung², and Darryl B Sneag^1
1Radiology, Hospital for Special Surgery, New York, NY, United States, ²GE Healthcare, Chicago, IL, United States

Advances in deep-learning algorithms aiming to improve image quality have not yet been well studied for their use in clinical interpretation. In this study, we compared interobserver agreement and image quality for lumbar spine (L-spine) MRI assessment of 3D T2-weighted fast spin echo (T2w-FSE) MRI, with and without deep learning (DLRecon) reconstructions, as well as standard-of-care (SOC) 2D T2w-FSE MRI. This pilot study demonstrated that interobserver agreement for variables of interest was good to very good regardless of reconstruction or sequence type, and overall image quality of DLRecon was not inferior despite significant reduction in scanning time.

Alejandro Morales Martinez¹, Jinhee Lee¹, Francesco Caliva¹, Claudia Iriondo¹, Sarthak Kamat¹, Sharmila Majumdar¹, and Valentina Pedoia^1
1UCSF, San Francisco, CA, United States

Large-scale analysis of the relationship between learned qMRI biomarkers and chronic knee pain. 7,437 patient timepoints reporting chronic pain were used to train three different deep learning models for bone shape, cartilage thickness, and cartilage T₂ biomarkers for the femur, tibia, and patella. The true chronic knee pain predictions for each trained model were further investigated with Grad-CAM and the max activation values for each model were sorted into clinically relevant anatomical compartments for each bone. Bone shape and cartilage T₂ seemed to be spatially correlated based on the results of the analysis. 

Sibaji Gaj¹, Dennis Chan¹, and Xiaojuan Li^1
1Department of Biomedical Engineering, Cleveland Clinic, Cleveland, OH, United States

Studies on systematic evaluations of effects of activation functions and loss functions on deep learning-based automated knee compartments segmentation models are limited. In this work, we present a 2D-UNet model for simultaneous automated bone and cartilage segmentation, and analyze the effect of different activation functions (rectified linear unit[relu], sigmoid and softmax) at all or last layer, and different loss functions (categorical cross-entropy, multiclass dice coefficient loss) with and without surface distance weights, on model performance. The results showed significant performance differences in average surface distance (ASD) between different activation functions. Adding surface distance to loss functions improved segmentation performances.

Martijn Froeling^1
1Imaging and oncology, University medical center utrecht, Utrecht, Netherlands

For accurate fitting of muscle diffusion tensor imaging data, many methods have been proposed. In this study, the performance of an unsupervised physics-informed deep learning method for IVIM-DTI fitting of muscle DTI data is investigated. The neural net comprised 9 fully connected networks and was tested on 20 upper leg DWI datasets. It trained in 45s and fitting a full dataset took around 4s. Although the parameter maps of the traditional and NN fitting look similar all parameters were significantly different. The network is capable of fitting the model within seconds but the differences need to be further investigated.

Andrew M Schmidt¹, Arjun D Desai¹, Lauren E Watkins², Hollis Crowder³, Elka B Rubin¹, Valentina Mazzoli¹, Quin Lu⁴, Marianne Black^1,3, Feliks Kogan¹, Garry E Gold^1,2, Brian A Hargreaves^1,5, and Akshay S Chaudhari^1,6
1Radiology, Stanford University, Stanford, CA, United States, ²Bioengineering, Stanford University, Stanford, CA, United States, ³Mechanical Engineering, Stanford University, Stanford, CA, United States, ⁴Philips Healthcare North America, Gainesville, FL, United States, ⁵Electrical Engineering, Stanford University, Stanford, CA, United States, ⁶Biomedical Data Science, Stanford University, Stanford, CA, United States

Automated segmentation using deep-learning can expedite segmentation tasks, but algorithm generalizability to unseen datasets is unknown. Here, we used two knee segmentation algorithms, each trained separately on Osteoarthritis Initiative double-echo steady-state (DESS) scans and quantitative DESS (qDESS) scans, to segment cartilage and meniscus from qDESS datasets from four independent studies. We compared manual-vs-automatic segmentation accuracy for morphology and T₂ map variations. We show that OAI-DESS-trained models may be suitable for quantifying relaxometry parameters in qDESS datasets but likely require fine-tuning to accurately quantify cartilage morphology. In contrast, qDESS-trained models generalize well to additional qDESS datasets for both morphology and T₂.

Sewon Kim¹, Hanbyol Jang², Seokjun Hong², Yeong Sang Hong³, Won C. Bae^4,5, Sungjun Kim*^3,6, and Dosik Hwang*^2
1Electrical and electronic engineering, Yonsei University, Seoul, Korea, Republic of, ²Yonsei University, Seoul, Korea, Republic of, ³Gangnam Severance Hospital, Seoul, Korea, Republic of, ⁴Department of Radiology, University of California-San Diego, San Diego, CA, United States, ⁵Department of Radiology, VA San Diego Healthcare System, San Diego, CA, United States, ⁶Yonsei University College of Medicine, Seoul, Korea, Republic of

We proposed a Bloch equation-based autoencoder regularization Generative Adversarial Network (BlochGAN) to generate T2-weighted fat saturation (T2 FS) images from T1-weighted (T1-w) and T2-weighted (T2-w) images for spine diagnosis. Our method can reduce the cost for acquiring multi-contrast images by reducing the number of contrasts to be scanned. BlochGAN properly generates the target contrast images by using GAN trained with the autoencoder regularization based on bloch equation, which is the basic principal of MR physics for identifying the physical basis of the contrasts. Our results demonstrate that BlochGAN achieved quantitatively and qualitatively superior performance compared to conventional methods.

Deepa Anand¹, Dattesh Shanbhag¹, Preetham Shankpal¹, Chitresh Bhushan², Desmond Teck Beng Yeo², Thomas K Foo², and Radhika Madhavan^2
1GE Healthcare, Bangalore, India, ²GE Global Research, Niskayuna, NY, United States

Data augmentation techniques have been routinely used in computer vision for simulating variations in input data and avoid overfitting. Here we propose a novel method to generate simulated images using features derived from activation maps of a deep neural network, which could mimic image variations due to MRI acquisition and hardware. Gradient-weighted Class Activation Mappings were used to identify regions important to classification output, and generate images with these regions obfuscated to mimic adversarial scenarios relevant for imaging variations. Training with images using the proposed data augmentation framework resulted in improved accuracy and enhanced robustness of knee MRI image classification.

Jihun Kwon¹, Masami Yoneyama¹, Takashige Yoshida², Kohei Yuda², Yuki Furukawa², Johannes M. Peeters³, and Marc Van Cauteren^3
1Philips Japan, Tokyo, Japan, ²Tokyo Metropolitan Police Hospital, Nakano, Japan, ³Philips Healthcare, Best, Netherlands

Shoulder MRI is typically acquired with multiple number of signals averaged (NSA) in order to average out breathing motion artifacts. However, higher NSA leads to a longer scan time and patient discomfort. In this study, we investigated the use of a deep learning-based reconstruction algorithm to highly accelerate shoulder MRI. Adaptive-CS-Net, a deep neural network previously introduced at the 2019 fastMRI challenge, was expanded and presented here as a Compressed-SENSE Artificial Intelligence (CS-AI) reconstruction. The purpose of this study was to compare the image quality of shoulder MRI between reference and accelerated methods; SENSE, Compressed-SENSE, and CS-AI.