Segmentation-based quantitative characterization of cervical disc degeneration on midsagittal T2-weighted MRI: a retrospective internal evaluation

To evaluate a semi-automated segmentation-based framework for quantitative assessment of cervical disc degeneration and to characterize the cross-sectional correspondence of signal- and geometry-derived indices with Pfirrmann imaging grades. In this retrospective single-center study, an experienced radiologist manually selected a midsagittal T2-weighted image, after which automated segmentation and quantitative analysis were performed within a semi-automated workflow. A two-dimensional Swin-UNETR model segmented vertebral bodies, intervertebral discs, cerebrospinal fluid, and the spinal cord. Segmentation performance was evaluated in an internal test cohort of 54 patients. Relative signal intensity (RSI), a Euclidean-distance-transform-derived height index (DH), and height-to-diameter ratio (HDR) were extracted from automated segmentations and compared with reference measurements in a spectrum-enriched nested subset of 34 patients comprising 170 cervical discs. A separate spectrum-enriched 30-patient subset underwent independent annotation and Pfirrmann grading to assess reader reliability. Associations between RSI, DH, HDR, and Pfirrmann grade were evaluated using patient-clustered analyses adjusted for age, sex, and cervical level. Five-class and simplified three-class grading models were additionally evaluated using leave-one-patient-out cross-validation. All analyses were internal; no external scanner or institutional cohort was evaluated. The Swin-UNETR model achieved Dice coefficients of 0.912 for vertebral bodies, 0.900 for intervertebral discs, and 0.821 for cerebrospinal fluid in the internal test cohort. Independent annotation showed strong inter-reader reliability, with an intervertebral-disc Dice of 0.873 (95% CI, 0.856–0.891) and linearly weighted Pfirrmann kappa of 0.901 (95% CI, 0.855–0.938). When the same EDT definition was applied to automated and reference masks on the physical scale, DH showed r = 0.894 and ICC(A,1) = 0.863; the stored historical series yielded ICC(A,1) values of 0.770 for RSI, 0.635 for DH, and 0.381 for HDR. RSI, stored DH, and stored HDR were inversely associated with Pfirrmann grade (Spearman’s ρ=−0.445, − 0.384, and − 0.635, respectively). After adjustment for age, sex, and cervical level, significant overall grade effects were observed for RSI ( P = 0.00164), DH ( P = 0.0185), and HDR ( P < 0.001). After Holm correction, significant adjacent-grade differences remained only for HDR between Grades III–IV and IV–V. Random Forest was the numerically highest-performing evaluated classifier, with accuracies of 0.488 and 0.665 for the five- and three-class tasks; the Decision Tree was retained for interpretability and achieved 0.435 and 0.635, respectively. The semi-automated framework enabled quantitative characterization of cervical disc degeneration through AI-assisted segmentation after manual midsagittal slice selection. Physical-scale DH showed strong automated-versus-reference agreement, while RSI and geometric indices demonstrated complementary cross-sectional correspondence with Pfirrmann grade. These associations represent correspondence with an imaging construct rather than independent biological or clinical validation. Classification findings were exploratory and internally evaluated only. Further multicenter and multi-scanner validation is required.