Computer-based measurements of joint space analysis using metacarpal morphometry in hand radiograph for evaluation of rheumatoid arthritis

Automated evaluation of rheumatoid arthritis from hand radiographs using Machine Learning and deep learning techniques

The aim and objectives of the study are as follows: (i) to implement automated patch-based classification of hand X-ray images using modified pre-trained convolutional neural network (CNN) models; (ii) to develop a customized CNN model for automated feature extraction and classification of hand X-ray images and to compare the performance of customized CNN models with non-linear and linear kernels; (iii) to construct the hand crafted feature fusion (SIFT+ Customized CNN features) and categorize the normal and RA using Machine Learning classifiers. The model was trained on 75 images (10,000 patches) of hand radiographs and tested using 25 images (500 patches) that were not included in the training set. The accuracy of the modified pre-trained model GoogLeNet was 89% and the proposed custom model three achieved an accuracy of 95%. The sensitivity and specificity of GoogLeNet were 84% and 90% respectively. The custom model three attained the sensitivity and specificity as 95% and 94% respectively. Furthermore, when compared to the features extracted (SIFT + CNN) from the customized models, the custom3 model outperformed well for the classification of RA compared to ML classifiers. Thus a custom CNN-based computer-aided diagnostic tool can be used as an effective method for the detection of RA.

Fully automated joint space width measurement and digital X-ray radiogrammetry in early RA

Objectives

To study fully automated digital joint space width (JSW) and bone mineral density (BMD) in relation to a conventional radiographic scoring method in early rheumatoid arthritis (eRA).

Methods

Radiographs scored by the modified Sharp van der Heijde score (SHS) in patients with eRA were acquired from the SWEdish FarmacOTherapy study. Fully automated JSW measurements of bilateral metacarpals 2, 3 and 4 were compared with the joint space narrowing (JSN) score in SHS. Multilevel mixed model statistics were applied to calculate the significance of the association between ΔJSW and ΔBMD over 1 year, and the JSW differences between damaged and undamaged joints as evaluated by the JSN.

Results

Based on 576 joints of 96 patients with eRA, a significant reduction from baseline to 1 year was observed in the JSW from 1.69 (±0.19) mm to 1.66 (±0.19) mm (p<0.01), and BMD from 0.583 (±0.068) g/cm² to 0.566 (±0.074) g/cm² (p<0.01). A significant positive association was observed between ΔJSW and ΔBMD over 1 year (p<0.0001). On an individual joint level, JSWs of undamaged (JSN=0) joints were wider than damaged (JSN>0) joints: 1.68 mm (95% CI 1.70 to 1.67) vs 1.54 mm (95% CI 1.63 to 1.46). Similarly the unadjusted multilevel model showed significant differences in JSW between undamaged (1.68 mm (95% CI 1.72 to 1.64)) and damaged joints (1.63 mm (95% CI 1.68 to 1.58)) (p=0.0048). This difference remained significant in the adjusted model: 1.66 mm (95% CI 1.70 to 1.61) vs 1.62 mm (95% CI 1.68 to 1.56) (p=0.042).

Conclusions

To measure the JSW with this fully automated digital tool may be useful as a quick and observer-independent application for evaluating cartilage damage in eRA.

Trial registration number

NCT00764725.

Validation of automatic joint space width measurements in hand radiographs in rheumatoid arthritis

Computerized methods promise quick, objective, and sensitive tools to quantify progression of radiological damage in rheumatoid arthritis (RA). Measurement of joint space width (JSW) in finger and wrist joints with these systems performed comparable to the Sharp-van der Heijde score (SHS). A next step toward clinical use, validation of precision and accuracy in hand joints with minimal damage, is described with a close scrutiny of sources of error. A recently developed system to measure metacarpophalangeal (MCP) and proximal interphalangeal (PIP) joints was validated in consecutive hand images of RA patients. To assess the impact of image acquisition, measurements on radiographs from a multicenter trial and from a recent prospective cohort in a single hospital were compared. Precision of the system was tested by comparing the joint space in mm in pairs of subsequent images with a short interval without progression of SHS. In case of incorrect measurements, the source of error was analyzed with a review by human experts. Accuracy was assessed by comparison with reported measurements with other systems. In the two series of radiographs, the system could automatically locate and measure 1003/1088 (92.2%) and 1143/1200 (95.3%) individual joints, respectively. In joints with a normal SHS, the average (SD) size of MCP joints was [Formula: see text] and [Formula: see text] in the two series of radiographs, and of PIP joints [Formula: see text] and [Formula: see text]. The difference in JSW between two serial radiographs with an interval of 6 to 12 months and unchanged SHS was [Formula: see text], indicating very good precision. Errors occurred more often in radiographs from the multicenter cohort than in a more recent series from a single hospital. Detailed analysis of the 55/1125 (4.9%) measurements that had a discrepant paired measurement revealed that variation in the process of image acquisition (exposure in 15% and repositioning in 57%) was a more frequent source of error than incorrect delineation by the software (25%). Various steps in the validation of an automated measurement system for JSW of MCP and PIP joints are described. The use of serial radiographs from different sources, with a short interval and limited damage, is helpful to detect sources of error. Image acquisition, in particular repositioning, is a dominant source of error.