Automatic MRI segmentation for upper limb muscles for clinical applications

点击翻译按钮获取中文摘要

SUMMARY The clinical value of magnetic resonance (MR) imaging of muscle anatomy and structure is limited by the bottleneck that arises from muscle segmentation. Thus, clinical evaluations of muscle in rotator cuff injuries to determine injury and repair potential use a single 2D image at a location selected primarily for consistency of anatomic landmarks; these methods are fundamentally flawed and do not accurately reflect either total volume or fatty infiltration for the rotator cuff muscles. Thus, effective, accurate, and fast methods for 3D segmentation for the upper limb are essential to enable the integration of valuable 3D imaging information into clinical decision- making. Our long-term goal is to develop a shareable framework that enables accurate, automated segmentation of MR images of upper limb muscles, on a timescale that makes image analysis tractable for the clinic. The overall objective is to leverage our fully annotated upper limb MR images in 48 healthy individuals and 10 persons with rotator cuff tears to develop, assess, and share successful machine learning approaches for both research and the clinic. Our central hypothesis is that supervised methods trained on our datasets will outperform unsupervised approaches and the resulting models can be successfully transferred to standard clinical scans. Our aims are to (1) identify the machine learning techniques with the best accuracy and performance for automatic segmentation of individual muscles in the upper limb from MR images, and (2) identify model generalizability and performance for analysis of parasagittal plane images. Our approach is to apply supervised techniques trained using our unique, existing, manually annotated images that include every muscle that crosses the shoulder, elbow, and wrist of 48 healthy individuals from three distinct age groups (25-35, 45-60, and 61-83 years) and the shoulder muscles of 10 elderly persons with rotator cuff tears. To strengthen potential translation to the clinical setting, we must consider application of these methods to the parasagittal plane in which clinical evaluation of muscle atrophy and fatty infiltration occurs; research scans are typically obtained in the axial plane. The expected outcomes are shareable models for segmentation of upper limb muscles and a computational framework to assess performance of a range of algorithms. Further, we expect to determine how effectively segmentation models, developed from our existing axial tomographic images of shoulder muscles, transfer for analysis of clinical images acquired to assess atrophy in rotator cuff injury. Accomplishing these objectives will provide the field the first set of open-source tools for automatic segmentation of upper limb muscles and will identify the critical next steps for enabling clinical translation.