Pancreatic Surface Lobularity in CT Imaging

• Pancreatic Surface Lobularity (PSL) is a quantitative imaging descriptor that measures the serrated anterior pancreatic margin to indicate early structural changes associated with T2DM.
• Automated segmentation using deep learning models and polynomial curve fitting robustly quantifies PSL, revealing statistically significant differences between diabetic and non-diabetic groups.
• Integrated within multivariate risk models, PSL serves as a sensitive biomarker for opportunistic CT screening by detecting early fibrotic and fatty alterations before overt morphological changes occur.

Pancreatic surface lobularity (PSL) is a quantitative imaging descriptor reflecting the serrated or “lobulated” morphology of the anterior pancreatic margin observed in cross-sectional imaging, particularly computed tomography (CT). Among patients with type-2 diabetes mellitus (T2DM), PSL increases as a consequence of chronic hyperglycemia, fatty infiltration, and micro-inflammation, resulting in structural changes such as acinar loss, interlobular fibrosis, and deepening of pancreatic incisures. PSL provides a sensitive biomarker for early structural pancreatic shifts and offers potential utility in opportunistic CT-based screening for T2DM, especially prior to overt volume loss or fat accumulation (Mathai et al., 13 Nov 2025).

1. Morphological Definition and Pathophysiological Context

PSL quantifies the deviation of the anterior pancreatic contour from a reference smooth curve, thereby formalizing the visual phenomenon of surface lobularity. In healthy individuals, the anterior pancreas border in axial CT slices is typically smooth. With progression of T2DM, the contour becomes increasingly serrated due to loss of acinar tissue and fibrotic evolution between lobules; this yields numerous small sinusoidal indentations. PSL is formally defined as the average (amplified) Euclidean distance between the true anterior surface and a least-squares polynomial approximation, evaluated across a standardized set of surface points. Clinically, increased PSL is correlated with subclinical and early fibrotic/fatty changes of the exocrine pancreas, making PSL sensitive to pre-diabetic states as well as manifest T2DM, and precedes the emergence of changes detectable by volumetry or fat fraction quantification (Mathai et al., 13 Nov 2025).

2. Automated Pancreatic Segmentation and PSL Quantification

A fully automated workflow was implemented to extract PSL from portal-venous phase abdominal CT, comprising organ segmentation, surface modeling, and PSL computation. Critical to this workflow is robust 3D pancreas segmentation. Four deep learning models were benchmarked:

• TotalSegmentator (TS): a public nnU-Net-based tool segmenting 104 anatomical structures.
• PanSegNet: an nnU-Net + Transformer hybrid trained on CT/MRI from multiple centers.
• Sub-Region nnU-Net: sequentially localizes head, body, tail landmarks and segments subregions.
• PancAP: a specialized 3D nnU-Net, trained on 1,350 PANORAMA CTs, utilizing anatomical priors to co-segment pancreas and vasculature.

Segmentation accuracy was assessed against expert-annotated CT references using the Dice similarity coefficient (DSC) and average symmetric surface distance (ASSD):

Model	Dice (mean ± SD)	ASSD (mm, mean ± SD)
TotalSegmentator	0.75 ± 0.16	2.19 ± 2.24
PanSegNet	0.78 ± 0.18	2.26 ± 3.12
Sub-Region	0.77 ± 0.17	2.05 ± 2.39
PancAP	0.79 ± 0.17	1.94 ± 2.63

PancAP exhibited statistically significant superiority over the competitors (p < 0.05, Wilcoxon with Bonferroni adjustment) (Mathai et al., 13 Nov 2025).

The PSL quantification pipeline involves:

1. Automatic extraction of the anterior pancreas surface points within seven contiguous axial slices centered on the largest cross-section.
2. Fitting a fourth-order polynomial curve $C(t)\;=\;\sum_{k=0}^4 a_k\,t^k$ to the anterior points via least squares.
3. Computing for each surface point $p_i$ the minimum distance $d_i = \min_{t} \|p_i - C(t)\|$.
4. Calculating slice-wise PSL as $$PSL_{\text{slice}} = 10\times\frac{1}{N}\sum_{i=1}^N d_i$$ (amplification factor).
5. Defining patient-level PSL as the median of the seven slice-wise scores.

3. Statistical Characterization of PSL in Diabetes

In the analyzed dataset (N = 584; 147 diabetic, 437 non-diabetic), quantification using the PancAP model revealed elevated PSL values for T2DM:

• Diabetic group: mean PSL = 4.26 ± 8.32
• Non-diabetic group: mean PSL = 3.19 ± 3.62
• Statistical significance: Wilcoxon rank-sum p = 0.01

The effect size as measured by Cliff’s delta was moderate, consistent with earlier semi-automated studies. These results support PSL as a biomarker that discriminates between diabetic and non-diabetic populations when derived from automated segmentation (Mathai et al., 13 Nov 2025).

4. PSL and Multivariate Risk Modeling for T2DM

To assess the clinical-dimensional utility of PSL, multivariate logistic regression ("glm" in R) was applied, partitioning the data into training (70%) and test (30%) cohorts. Three modeling paradigms were evaluated:

• Clinical only: age, body-mass index (BMI)
• Imaging only: a panel of 16 CT biomarkers, including PSL, organ and fat volumes, and mean/SD of multiple tissue attenuation metrics
• Combined: clinical and imaging predictors

The logistic regression model was:

$$\log\frac{P(\mathrm{T2DM}=1)}{1 - P(\mathrm{T2DM}=1)} = \beta_0 + \sum_{j}\beta_j\,x_j$$

where $x_j$ includes PSL and other standardized features.

On held-out test data, PancAP-based pipeline performance was:

Model	AUC (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)
Clinical only	0.75 (0.65, 0.84)	35.9% (20.5, 51.3)	93.4% (88.3, 97.1)
Imaging only	0.89 (0.81, 0.96)*	74.3% (58.9, 87.2)*	89.8% (84.7, 94.2)
Combined	0.90 (0.83, 0.96)*	66.7% (51.3, 79.6)	91.9% (87.6, 96.4)

(*p < 0.01 vs. clinical only by bootstrap ROC test; imaging vs. combined not significantly different.) High specificity at the Youden threshold indicates strong negative predictive value for opportunistic screening. Calibration and Brier scores were not reported (Mathai et al., 13 Nov 2025).

5. Clinical Utility, Workflow Integration, and Limitations

PSL can be integrated into a secondary analysis pipeline for any patient undergoing abdominal CT (irrespective of indication), enabling automated pancreas segmentation and calculation of PSL with other CT biomarkers. Elevated PSL or composite risk above a prespecified threshold can trigger referral for confirmatory laboratory testing, supporting earlier detection of pre-diabetes or T2DM.

PSL demonstrates the potential to detect early pancreatic dysfunction before gross morphologic alterations and, by extension, may serve as an indicator for intervention prior to irreversible damage. Limitations cited include lack of sex adjustment (noting men exhibit higher lobularity), dependency on high-quality segmentation (though PancAP failed in only 0.17% of scans), use of only whole-pancreas rather than sub-regional PSL, and retrospective, single-center paper constraints; multi-center external validation remains outstanding. Other limitations include potential mis-segmentation errors propagating into PSL quantification, and lack of calibration analysis. Extension to non-contrast CT and MRI, incorporation of sex-specific norms, and application of PSL to other pancreatic diseases constitute areas for future research (Mathai et al., 13 Nov 2025).

6. Future Directions and Research Implications

The objective quantification of PSL via automated segmentation and polynomial-based surface modeling represents a scalable paradigm for radiomics in metabolic and exocrine pancreatic disease. Planned directions include non-contrast CT and MRI adaptation, regional PSL analysis, and assessment for other endocrine and exocrine pancreatic conditions such as chronic pancreatitis. Multicenter validation and the derivation of sex-specific normative metrics are needed to establish population-wide reference intervals and optimize stratification for early screening. The embedding of PSL in multivariate CT-based risk models achieving AUC ≈ 0.90 indicates its robust discriminative potential for asymptomatic T2DM opportunistic identification (Mathai et al., 13 Nov 2025).