A Deep Registration Method for Accurate Quantification of Joint Space Narrowing Progression in Rheumatoid Arthritis

TL;DR

This study targets accurate, quantitative monitoring of joint-space narrowing (JSN) progression in early rheumatoid arthritis (RA) from radiographs. It introduces a two-network framework: a U-net++-based joint segmentation module and a ResNet-like intra-subject rigid registration network that estimates separate four-parameter transformations for the upper and lower joint regions, yielding JSN via $JSN_{fg} = dy_0 - dy_1$. The approach delivers sub-pixel accuracy, robustness to rotation, scaling, and noise, and provides a visualization loss to assess reliability, demonstrated on a clinical hand radiograph dataset with mean-squared error $0.0031$, standard deviation $0.0661\text{ mm}$, and mismatching rate $0.48\%$. This method offers improved sensitivity for JSN progression over existing techniques, with potential for clinical deployment in monitoring drug responses and disease progression in RA. The work also reports strong segmentation performance (high mIoU and DSC) and robust JSN quantification across finger joints, supporting its applicability in routine radiographic assessment.

Abstract

Rheumatoid arthritis (RA) is a chronic autoimmune inflammatory disease that results in progressive articular destruction and severe disability. Joint space narrowing (JSN) progression has been regarded as an important indicator for RA progression and has received sustained attention. In the diagnosis and monitoring of RA, radiology plays a crucial role to monitor joint space. A new framework for monitoring joint space by quantifying JSN progression through image registration in radiographic images has been developed. This framework offers the advantage of high accuracy, however, challenges do exist in reducing mismatches and improving reliability. In this work, a deep intra-subject rigid registration network is proposed to automatically quantify JSN progression in the early stage of RA. In our experiments, the mean-square error of Euclidean distance between moving and fixed image is 0.0031, standard deviation is 0.0661 mm, and the mismatching rate is 0.48\%. The proposed method has sub-pixel level accuracy, exceeding manual measurements by far, and is equipped with immune to noise, rotation, and scaling of joints. Moreover, this work provides loss visualization, which can aid radiologists and rheumatologists in assessing quantification reliability, with important implications for possible future clinical applications. As a result, we are optimistic that this proposed work will make a significant contribution to the automatic quantification of JSN progression in RA.

Figures (9)

• Figure 1: JSN progression of a MCP joint for little finger over a period of 10 months. From left to right the images are: baseline, five-month, and ten-month images (spatial resolution: 0.175 mm/pixel). Usually, JSN progression is less than one pixel per year, therefore, it is difficult for radiologists and rheumatologists to see. Operating with an algorithm with pixel level accuracy to quantify JSN progression over a period of one year can be ineffective. JSN progression measured for five and ten months X-rays relative to baseline using this work are -0.197 pixel and 0.174 pixel respectively.
• Figure 2: (A) Four rigid transformation parameters are shown that are used in this work; $dz$: scaling, $d\theta$: rotation, $dx$: displacement on x-axis, $dy$: displacement on y-axis. (B) The overview of our proposed deep learning image registration based JSN progression quantification methodology. This work can be divided into two steps: joint segmentation, and JSN progression quantization. Take a MCP joint as an example, this work can be performed as follow: (i) § \ref{['sec:segmentation']} A supervised U-net++ based network is implemented to segment the proximal phalanx bone and metacarpal bone region of the MCP joint. (ii) § \ref{['sec:JSNQuanti']} An un-supervised ResNet-like based deep registration network is proposed to quantify the rigid transformation parameters of the proximal phalanx bone and metacarpal bone region. (iii) The JSN progression can be obtained by calculating the displacement difference on y-axis between two bone region.
• Figure 3: The diagram of our segmentation network. This segmentation network contains one convolutional layer (kernel size: $7\times7$, channels: 64) and a 5-layer Unet++ network.
• Figure 4: The structure diagram of image registration network. In this case, after a convolution layer of $3\times3$ convolution kernels and a 1-channel convolution base and combined with its corresponding segmentation mask, the input image set is input into the registration network. The registration network contains a layer of convolution and 4 layers of residual convolution modules with the channels of 64, 128, 256, 512 respectively. The final transformation parameters are obtained after a full connection layer with 8 channels as output. These transformation parameters are used to deform the moving image to generate a warped image, and the difference between the generated warped image and the fixed image is defined as the loss, which is used to optimize the registration network.
• Figure 5: Experiments of the proposed segmentation network. White lines represent the manual label of segmentation, and yellow lines represent the predicted segmentation by using the network.