Featured Investigators


Valentina Pedoia, PhD and Drew Lansdown, MD - Multi-Task Deep Learning to Develop Automatic Scrapular Shape Extraction from Clinical MR

This study aims to use deep learning to perform automatic scapula bone segmentation and to simultaneously synthetize CT-like images. With funding from the CCMBM Pilot and Feasibility Grant Program, the proposed study builds a novel translational platform to revolutionize shoulder MR images in research studies, but also is paradigm-shifting in that it may provide a first step towards a more quantitative approach to surgical planning and patient management.

Figure 1

Figure 2

Methodology Needed to Obtain Bony and Soft Tissue Information
Scapular bone shape is an important determinant in surgical planning and predictor of post-operative outcomes for patients with shoulder instability and shoulder osteoarthritis. Currently, clinical evaluation of scapular bone shape is performed on a three-dimensional (3D) computed tomography (CT) scan, while a magnetic resonance imaging (MRI) scan is obtained to evaluate the soft tissue surrounding the shoulder. There is a clear clinical and research need for a methodology to obtain bony and soft tissue information from a single imaging study in an accurate, repeatable and fully automated fashion. 

Deep Learning Allows Automated Segmentation of Cartilage and Bone

Deep Learning (DL), especially convolutional neural networks (CNNs), has made strides in several domains as speech recognition, visual object detection, classification, drug discovery and genomics. DL shines when afforded large datasets, as its automated feature extraction allows one to solve problems too complex for conventional approaches. CNNs are representation learning methods characterized by the usage of multiple, simple, but non-linear units to build several interconnected layers. Each layer aggregates the information at increasing levels of abstraction starting with simple image elements, as edges or contrast, to more complex and semantic aggregations, uncovering latent patterns able to accomplish pattern recognition tasks. DL has allowed for the automatic segmentation of knee cartilage, knee and hip bone; however, the scapula is a unique thin structure with a complex shape that poses different challenges.  

Novel Directions for Shoulder Instability and Should Osteoarthritis in Clinical Treatment and Research

We have performed a pilot study utilizing a two-dimensional (2D) V-net convolutional neural network architecture with good results (mean Dice score coefficient: 82%), however further improvement is necessary for clinical or research implementation. We propose applying mixed precision training to allow for 3D processing in an efficient fashion. We also plan to augment our dataset with Statistical Shape Modeling to generate multiple synthetic training examples to be used for model pre-training and transfer learning. Finally, we plan to apply multi-task learning to simultaneously synthesize CT images and segment scapular bone from a standard clinical MRI scan. We will compare model performance on the MRI scan relative to a matched 3D-CT scan from previously acquired patient scans. The results of this study have the potential to greatly impact clinical treatment and create novel directions of research on shoulder instability and shoulder osteoarthritis.


Headshot of Valentina Pedoia, PhD

Valentina Pedoia, PhD 

Assistant Professor

Department of Radiology and Biomedical Imaging


Research interests: Medical imaging, computer vision, machine learning, big data analysis, MRI, musculoskeletal imaging, clinically oriented quantitative imaging, articular cartilage compositional imaging

Headshot of Drew Lansdown, MD

Drew Lansdown, MD

Assistant Professor

Department of Orthopaedic Surgery

Research interests: Musculoskeletal quantitative imaging, sports medicine, ligament imaging, muscle imaging



Back to Top^