Automatic spondylolisthesis grading from MRIs across modalities using faster adversarial recognition network.

Abstract

Grading spondylolisthesis into several stages from MRI images is challenging because detecting critical vertebrae and locating landmarks in images of different characteristics is difficult. We propose Faster Adversarial Recognition (FAR) network to accurately perform spondylolisthesis grading by excellently detecting critical vertebrae without the need of locating the landmarks. The FAR network introduces the adversarial scheme by using a multi-task recognition network as the generator and an adversarial module as the discriminator. The multi-task recognition network (generator) is an integrated network that can reliably perform multi-scale hierarchical feature learning, critical vertebrae detection, detected vertebrae classification, bounding box regression, and spondylolisthesis grading in a hybrid supervised manner. The adversarial module (discriminator) takes the detection results as inputs to supervise the generative network by leveraging the high-order statistics of the distribution of the detected bounding box coordinates. The FAR network is evaluated to be accurate and robust in spondylolisthesis grading (training accuracy: 0.9883 ± 0.0094, testing accuracy: 0.8933 ± 0.0276) for MRI images of different modalities, which can be attributed to the excellent critical vertebrae detection (detection mAP75 for training: 1 ± 0, for testing: 0.9636 ± 0.0180, and IoU (Intersection-over-union) ≥ 0.9/0.8 for most detections with their corresponding ground truth in the training/testing dataset). This accuracy is comparable to that of the physicians and outperforms other state-of-the-art methods. These results indicate the potential of our framework to perform spondylolisthesis grading for clinical diagnosis.