Multi-Step Segmentation Algorithm for Quantitative Magnetic Resonance Imaging T2 Mapping of Ruptured Achilles Tendons

Abstract

Diagnosis and treatment monitoring of Achilles tendon ruptures (ATRs) are supported by medical imaging, in particular by magnetic resonance imaging (MRI) for tendon volume and healing assessment. Therefore, we propose an automatic, multi-step segmentation algorithm for quantitative MRI T2 mapping of ATRs. Seventy retrospective post-trauma, post-surgery and follow-up studies were included in this research. The automatic segmentation algorithm for the inhomogeneous, noisy Achilles tendon region consisted of a multi-step anisotropic denoising, T2 map reconstruction with a weighted log-linear regression, thresholding with T2 time parameters, region growing and morphological closing. The automatic segmentation results were compared with those from manual contour tracing (MCT) performed by two radiologists. The Intersection over Union (IoU), specificity, sensitivity, F1-score, Yasnoff's normalized distance (YND), and type I and II errors were used to assess the segmentation accuracy. The segmentation methods were also compared with a Bland-Altman plot of the volumes of the segmented regions, with mean differences, correlation coefficients and 95% confidence intervals. The mean specificity and sensitivity values were high, 99.8± 0.1% and 85.9± 8.7%, respectively, with corresponding type I and II errors of 0.2± 0.1% and 14.1± 8.7%. The IoU, F1-score and YND were 71.0± 9.2%, 82.7± 6.3% and 0.007± 0.007%, respectively. The tendon volumes obtained by manual and automatic segmentation were strongly positively correlated ($\text{R}^{2} =0.85$ ), and the Bland-Altman plot depicted good comparability. The average difference was -28 voxels (95% confidence interval: -2726 to 2782 voxels). For ATRs, our method is reliable, with a strong positive correlation with MCT and a very high specificity.