Automatic segmentation and grading of ankylosing spondylitis on MR images via lightweight hybrid multi-scale convolutional neural network with reinforcement learning

Objective.Ankylosing spondylitis (AS) is a disabling systemic disease that seriously threatens the patient's quality of life. Magnetic resonance imaging (MRI) is highly preferred in clinical diagnosis due to its high contrast and tissue resolution. However, since the uncertainty and intensity inhomogeneous of the AS lesions in MRI, it is still challenging and time-consuming for doctors to quantify the lesions to determine the grade of the patient's condition. Thus, an automatic AS grading method is presented in this study, which integrates the lesion segmentation and grading in a pipeline.Approach. To tackle the large variations in lesion shapes, sizes, and intensity distributions, a lightweight hybrid multi-scale convolutional neural network with reinforcement learning (LHR-Net) is proposed for the AS lesion segmentation. Specifically, the proposed LHR-Net is equipped with the newly proposed hybrid multi-scale module, which consists of multiply convolution layers with different kernel sizes and dilation rates for extracting sufficient multi-scale features. Additionally, a reinforcement learning-based data augmentation module is utilized to deal with the subjects with diffuse and fuzzy lesions that are difficult to segment. Furthermore, to resolve the incomplete segmentation results caused by the inhomogeneous intensity distributions of the AS lesions in MR images, a voxel constraint strategy is proposed to weigh the training voxel labels in the lesion regions. With the accurately segmented AS lesions, automatic AS grading is then performed by a ResNet-50-based classification network.Main results. The performance of the proposed LHR-Net was extensively evaluated on a clinically collected AS MRI dataset, which includes 100 subjects. Dice similarity coefficient (DSC), average surface distance, Hausdorff Distance at95thpercentile (HD95), predicted positive volume, and sensitivity were employed to quantitatively evaluate the segmentation results. The average DSC of the proposed LHR-Net on the AS dataset reached 0.71 on the test set, which outperforms the other state-of-the-art segmentation method by 0.04.Significance. With the accurately segmented lesions, 31 subjects in the test set (38 subjects) were correctly graded, which demonstrates that the proposed LHR-Net might provide a potential automatic method for reproducible computer-assisted diagnosis of AS grading.