Automatic Detection of Increased Signal Intensity from T2-Weighted Magnetic Resonance Images (MRIs) of Cervical Vertebra

Abstract

Automatic detection of increased signal intensity (ISI) in spinal cord from T2-weighted magnetic resonance images (MRIs) of the cervical vertebra was design. To diagnose the conditions of spinal cord automatically and aid doctors in treating patients with cervical degenerative diseases. 724 MR T2 images were annotated by two doctors with ISI information and 245 MR T2 ones with the mask of cervical discs and spinal cord. The proposed approach contains three modules: (1) a module named region proposal to generate the ISI proposals. This module is based on three submodules: (a) a Mask R-CNN model to segment the cervical disc and spinal cord, (b) a rule-based algorithm for verifying the segmentation result, (c) an algorithm for getting the segment of spinal cord according to the verified segmentation result; (2) a feature extraction method based on Laplace operator to extract the feature of segments of spinal cord; (3) a classifier based on Linear Discriminant Analysis to classify the segments of spinal cord. Our ISI detection method achieved 31.9 AP in the test. And while using Fβ score (β = 1.3) the method got 85.4% recall and 2.16 FP per image. In this study, we proposed an approach based on Mask R-CNN to detect ISI on MR T2 images automatically and achieved good performance both in accuracy and speed.