Deep Learning for Detection of Structural Sacroiliac Joint Lesions on Pelvic Computed Tomography: Multicenter Development and Validation

Purpose or Learning Objective: To evaluate the feasibility and diagnostic accuracy of deep learning for automated detection of structural lesions of sacroiliitis on multicenter pelvic computed tomography (CT).

Methods or Background: A heterogeneous multi-scanner pelvic dual-energy computed tomography data set of 145 patients (81 women; 121 Center A/24 Center B; 18–87 years of age, mean 40 ± 13 years; 2005–2021) with a clinical suspicion of sacroiliitis was included retrospectively.

Ground truth manual pixelwise sacroiliac joint (SIJ) segmentation and structural lesion annotation was performed on axial CT images by three independent pretrained radiologists separately, blinded for clinical information. Erosions were defined as a cortical bone full-thickness loss ≥ 1.0 mm. Ankylosis was defined as SIJ bridging ≥ 2.0 mm. The reference standard was the reader-assessed presence of structural SIJ lesions.

Preprocessing steps were performed to homogenize the heterogeneous original images from different scanners to conform to identical image quality and thus improve statistical outcome performance. A U-Net for SIJ segmentation and two separate convolutional neural networks (CNNs) for erosion and ankylosis detection were trained. In-training validation and 10-fold validation testing (U-Net-n = 10 × 58; CNN-n = 10 × 29) on a separate test data set were conducted to assess performance on a slice-by-slice and patient level (Dice coefficient/accuracy/sensitivity/specificity/positive and negative predictive value/receiver operating characteristic-area under the curve [ROC-AUC]). Patient-level optimization was applied to increase the performance regarding predefined statistical metrics.

Gradient-weighted class activation mapping (Grad-CAM++) heat map explainability analysis highlighted image parts with statistically important regions for algorithmic decisions.

Results or Findings: A total of 84 patients had a diagnosis of spondyloarthritis, 15 had a mechanical low back and/or buttock pain origin, and 46 had no clear diagnosis. Sixty-four patients had SIJ erosions, 28 had ankylosis, 14 had both, and 67 had no structural lesions.

Regarding SIJ segmentation, a training and test Dice similarity coefficient of 0.89 and 0.75, respectively, were obtained. By segmenting the SIJs before the disease detection steps, the CTs were reduced to a third of the total number of slices of the original CT. For slice-by-slice structural lesion detection, a sensitivity/specificity/ROC-AUC of 95%/89%/0.92 and 93%/91%/0.91 were obtained in the test data set for erosion and ankylosis detection, respectively. For patient-level lesion detection after pipeline optimization for predefined statistical metrics, a sensitivity/specificity of 95%/85% and 82%/97% were obtained for erosion and ankylosis detection, respectively.

Grad-CAM++ explainability analysis showed that the focus of the CNN to detect structural lesions is on the SIJ cortical edges, as intended.

Conclusion: An optimized deep learning pipeline, including an explainability analysis, detects structural lesions of sacroiliitis on pelvic CTs with excellent statistical performance on a slice-by-slice and patient level.

Publication History

Article published online:
26 May 2023