Using the Deep Learning Algorithm to Determine the Presence of Sacroiliitis from Pelvic Radiographs

Abstract

Deep learning (DL) techniques have demonstrated remarkable capabilities in recognizing complex patterns in medical imaging data. In recent years, DL has been increasingly applied in clinical medicine for disease diagnosis and progression prediction. This study aimed to develop and validate a DL model for detecting sacroiliitis using pelvic anteroposterior (AP) radiographs. We retrospectively analyzed 1853 patients with pelvic AP radiographs, including 3706 sacroiliac joints (SIJs). Pelvic AP radiographs served as input data for the DL model development, while the presence or absence of sacroiliitis confirmed by pelvic computed tomography (CT) was used as the reference standard output data. Based on CT findings, 1463 of 1853 right SIJs showed evidence of sacroiliitis, while 390 had no sacroiliitis. Similar findings were observed in the left SIJs. The dataset was split with 70% (1297 images) for training and 30% (556 images) for validation. The areas under the curve (AUC) for our DL model on the validation dataset were 0.871 (95% confidence interval (CI): 0.834-0.907) and 0.869 (95% CI: 0.834-0.907) for the left and right SIJs, respectively. Diagnostic accuracies for sacroiliitis on the left and right sides were 85.4% and 86.3%, respectively. These results demonstrate that a DL model trained on pelvic AP radiographs with CT-confirmed diagnoses can effectively aid in the diagnosis of sacroiliitis.

Keywords: computed tomography; convolutional neural network; deep learning; radiograph; sacroiliac joint; sacroiliitis.

Figure 1

(A) Absence of sacroiliitis on bilateral sacroiliac joints. (B) Presence of sacroiliitis on bilateral sacroiliac joints.

Figure 2

Flowchart illustrating the deep learning process for detecting sacroiliitis in pelvic anteroposterior radiographs, from image acquisition through anteroposterior radiographs and from image acquisition through region of interest extraction to final classification. Abbreviations: AP—anteroposterior; ROI—region of interest; AUC—area under the curve.

Figure 3

Region of interest and heatmap images. The heatmap colors delineate the intensity of model attention or activation within the region of interest. Warmer colors (red and orange) highlight areas of high importance that significantly influenced the model’s prediction. Cooler colors, (blue and purple) indicate regions of lower importance with minimal impact on the outcome.

Figure 4

Receiver operating characteristic curve for the validation and test dataset of the anteroposterior pelvic radiographs for the deep learning model for diagnosing sacroiliitis. SIJ—sacroiliac joint; Acc—accuracy; AUC—area under the curve; CI—confidence interval.

Figure 5

Confusion matrix for our developed models (tested with the validation dataset) (0—absence of sacroiliitis; 1—presence of sacroiliitis). SIJ—sacroiliac joint.