Knee articular cartilage morphometrics is an effective tool to derive imaging biomarkers for knee osteoarthritis (OA) (Gray et al., 2004, Mosher et al., 2011, Hunter et al., 2014, Eckstein et al., 2015, Collins et al., 2016, Guermazi et al., 2017, Wirth et al., 2017, Everhart et al., 2019, Hunter et al., 2022). Magnetic resonance (MR) imaging is a predominant non-invasive imaging modality for OA research (Oei et al., 2021) because of its ability to anatomically and biochemically quantify alterations in the cartilage at different stages of OA progression. From a structural MR image, morphological biomarkers can be analyzed and correlated with clinical outcomes. Tremendous efforts have been made to develop automated analysis techniques for imaging biomarker extraction. On the one hand, semiautomated methods for the quantification of cartilage lesions, thickness, surface area, and volume have been proposed (Cohen et al., 1999, Stammberger et al., 1999, Hohe et al., 2002, Kauffmann et al., 2003, Eckstein et al., 2006b, Eckstein et al., 2006c, Carballido-Gamio et al., 2008, Wirth and Eckstein, 2008, Williams et al., 2010, Maerz et al., 2016, Favre et al., 2017). On the other hand, semiquantitative methods for the grading of cartilage lesions have also been proposed (Eckstein et al., 2006b, Guermazi et al., 2017, Dório et al., 2020). These semiquantitative methods usually rely on MR-based grading systems, such as WORMS (Peterfy et al., 2004), KOSS (Kornaat et al., 2005), BLOKS (Hunter et al., 2008), MOAKS (Hunter et al., 2011), and CaLS (Alizai et al., 2014). Semiautomated and semiquantitative methods are subjective, error-prone, and time-consuming as they require human interaction and domain knowledge. To overcome such drawbacks, automated tissue segmentation methods have been extensively exploited and validated (Prasoon et al., 2013, Raj et al., 2018, Liu et al., 2018, Zhou et al., 2018, Ambellan et al., 2019, Tan et al., 2019, Xu and Niethammer, 2019, Gaj et al., 2020, Khan et al., 2022). To the best of our knowledge, there is a lack of automated methods for the quantification of cartilage lesions. To fill this gap, we aimed to automate the quantitative measurement of full-thickness cartilage loss (FCL) which has been shown to be an imaging biomarker for OA progression (Guermazi et al., 2017, Everhart et al., 2019). Additionally, we developed methods for robust cartilage thickness mapping, rule-based cartilage parcellation, and regional quantification.

In this work, we developed a framework called CartiMorph for automated knee articular cartilage morphometrics. The quantitative

metrics for cartilage subregions include the percentage of FCL, mean thickness, surface area, and volume. Deep learning models were integrated into the proposed framework for tissue segmentation, template construction, and template-to-image registration. Our code and experimental details are publicly available.1 The contributions of this work are summarized as follows.

We established a method for automated cartilage thickness mapping that is robust to cartilage lesions.

We established a method for automated FCL estimation through learning-based deformable image registration, template construction, and surface-based operations.

We established a rule-based cartilage parcellation method that automatically partitions femoral and tibial cartilages into 20 subregions. The parcellation method was designed for robust regional quantification despite the extent of the FCL.

We constructed a knee template image learned from MR scans of subjects without radiographic OA. We additionally constructed the respective 5-class tissue segmentation and a 20-region atlas for the template image, which provide prior knowledge of the bone and cartilage anatomy.

We obtained new insights into the effectiveness of deep-learning-based segmentation models in cartilage morphometrics.

The remainder of this article is organized as follows. Section 2 describes related works on tissue segmentation 2.1, image registration 2.2, template construction 2.3, and cartilage morphometrics 2.4. Section 3 provides details of the proposed framework. Section 4 presents the experiments and results. Section 5 discusses the insights gained from the results, the limitations of this work, and future works. We draw conclusions from the results in Section 6.