Artificial Intelligence‐Estimated Intrapancreatic Fat and Its Relationship to Familial and Germline Pancreatic Cancer Risk Authors: Tom Konikoff, Laurens P Janssens, Hiroaki Takahashi, Garima Suman, Fred Nugen, Anup Mishra, Panos Korfiatis, Erin E Carlson, William R Bamlet, Aishwarya Pradeep, Kelly Harper, Karen A Doering, Adriana Delgado, Bradley J Erickson, Ajit H Goenka, Naoki Takahashi, Ann L Oberg, Shounak Majumder Published: 2026 ABSTRACT ABSTRACTBackgroundFamily history of pancreatic cancer (PC), pathogenic germline variants (PGVs), and increased intrapancreatic fat (IPF) are individually associated with increased PC risk. Whether IPF is elevated in high‐risk individuals (HRIs) with PC family history or PGVs is unknown. Although MRI is the preferred modality for IPF measurement, region‐of‐interest–based approaches are prone to sampling bias because pancreatic fat is heterogeneously distributed. We aimed to compare IPF among HRIs and matched non‐HRI controls using an Artificial intelligence (AI)‐based tool for whole pancreatic IPF measurement applied to magnetic resonance imaging (MRI).MethodsWe developed and validated a deep‐learning pipeline for IPF measurement that incorporates automated pancreas segmentation on T1‐weighted MRI images registered to the corresponding iterative decomposition of water and fat with echo asymmetry and least squares estimation images to compute IPF using proton density fat fraction (PDFF). This pipeline was used to quantify IPF and assess IPF distribution patterns in 85 HRIs and 170 non‐HRIs matched 1:2 for age, sex, BMI, and diabetes status. A surface erosion approach was developed to minimize the impact of peripancreatic fat on IPF measurement. Linear models adjusting for matching variables were used to assess associations.ResultsModel‐expert agreement for whole‐pancreas PDFF was excellent (intraclass correlation coefficient 0.99; 95% CI, 0.98–0.99 mean absolute error 1.4%). Mean PDFF did not differ between HRIs and controls (25.8% [11.5] vs. 26.9% [11.4]; p = 0.32). PDFF increased with age (p = 0.015), male sex (p < 0.001), presence of diabetes (p = 0.009), BMI (p < 0.001). IPF distribution patterns did not differ by risk group.ConclusionAn AI‐based approach provides accurate, scalable, and objective assessment of IPF quantity and distribution. IPF quantity and spatial distribution do not appear to differ based on germline or familial predisposition to PC.Keywords: artificial intelligence, fatty pancreas, intrapancreatic fat, pancreatic cancer, pancreatic imaging Abbreviations high risk individuals iterative decomposition of water and fat with echo asymmetry and least squares estimation intrapancreatic fat pancreatic cancer proton density fat fraction Introduction Intrapancreatic fat (IPF) has been increasingly recognized for its association with various metabolic disorders, including diabetes mellitus (DM), obesity, and pancreatic diseases, including pancreatitis and pancreatic cancer (PC) [1, 2, 3, 4, 5, 6]. Magnetic resonance imaging (MRI) is generally considered a gold standard for measuring IPF and correlates with histology [4, 7, 8]. Nevertheless, no standardized objective estimation method and diagnostic thresholds for IPF exist, and in current clinical practice, IPF is estimated manually from MRI sequences by radiologists either subjectively based on image review or by selecting multiple operator‐defined regions of interest (ROIs) on the pancreas and calculating the average fat content to determine the overall IPF. Compared to fat in other organs such as the liver, IPF is heterogeneously distributed, and hence this ROI‐based approach increases the risk of sampling error [9, 10, 11, 12]. The use of ROIs has yielded conflicting results in the literature and represents a fundamental barrier to establishing IPF as a reliable biomarker. With advances in Artificial Intelligence (AI), recent studies have shown increasing use of automated segmentation of the pancreas for IPF estimation [13]. To enable accurate scalable quantification of IPF and facilitate objective studies focused on its relationship with various pancreatic disorders, our group and others have led efforts focused on developing non‐invasive and accurate AI‐enabled methods, specifically using an image‐based whole‐organ approach [9, 14, 15, 16, 17]. PC is a lethal disease and current early detection efforts focus on image‐based screening using MRI and endoscopic ultrasound in high‐risk individuals (HRIs), defined as those with pathogenic or likely pathogenic germline variants in PC susceptibility genes and/or a family history of PC affecting two or more blood relatives [18, 19]. While the exact role IPF plays in PC is not well understood in humans, there are emerging data suggesting IPF driving PC [13]. In genome‐wide association studies, genetic variants linked to increased IPF have been associated with an increased susceptibility to pancreatic ductal adenocarcinoma (PDAC) and this association remained significant even after excluding genetic variants related to body mass index (BMI) [20]. Given this emerging evidence linking genetics and IPF to pancreatic carcinogenesis, we hypothesized that IPF accumulation may also be a phenotypic marker of germline or familial predisposition to the disease. However, this has never been systematically investigated [6, 13, 21, 22, 23]. There are no studies reporting IPF estimates in HRIs and it is not known if HRIs exhibit higher IPF compared to non‐HRIs. Furthermore, the presence of increased IPF may also reduce the contrast between normal tissue and small tumors, making early detection, which is specifically important in HRIs, more difficult. In this two‐step study, we aimed first to develop and validate a robust AI pipeline for whole‐organ IPF quantification from clinical iterative decomposition of water and fat with echo asymmetry and least squares estimation (IDEAL) MRI sequences using mean proton density fat fraction (PDFF), overcoming the challenges of manual, ROI‐based methods. Then, we applied this novel tool to definitively test the association between IPF quantity and spatial distribution and familial/genetic risk for PC in rigorously matched case‐control cohorts. Methods This Mayo Clinic Institutional Review Board approved study included two distinct steps. The first step focused on the development of an AI‐based model for whole organ IPF quantification (IRB # 16‐007866). The second step utilized the AI model to estimate and compare IPF between HRIs and non‐HRI controls (IRB # 22‐000996) in a matched case‐control study. The study follows The Strengthening the Reporting of Observational Studies on Epidemiology (STROBE) statement [24]. The study workflow is outlined in Figure 1. Patients were not actively involved in the study design process. Study flow diagram: Step 1: development and validation of a convolutional neural‐network (nnUNet) to segment the pancreas on T1‐weighted sequences and register the mask to iterative decomposition of water and fat with echo asymmetry and least squares estimation (IDEAL) FatFrac sequences for whole‐pancreas proton‐density‐fat‐fraction (PDFF) quantification. Step 2: application of the trained model to a cohort of high‐risk individuals (HRIs) and matched non‐HRI controls to compare intrapancreatic fat (IPF), perform embedding‐based clustering, and multivariable statistical analyses. *Chosen out of the original abdominal IDEAL database (n = 1940), excluding studies used for developing the model. BMI = body mass index; DM = diabetes mellitus; HRPC = high‐risk pancreatic clinic; PCA = principal component analysis. Step 1: AI Model Development for IPF Quantification Patients with abdominal MRI examinations, including T1 weighted post‐contrast and IDEAL sequences, aged 40–80 years at the time of imaging and that had provided Minnesota research authorization, were identified at Mayo Clinic, Rochester between January 2020 and December 2022, yielding a cohort of 1940 patients. MRIs were performed with 3.0‐T and 1.5‐T scanners (DISCOVERY MR750w, Optima MR450w, and SIGNA Premier; GE MEDICAL SYSTEMS, U.S.A.). MRI technique parameters for all cohorts can be found in Supporting Information S1: Tables S1 and S2. The IDEAL sequence is considered the most accurate sequence for assessing IPF and is quantified as proton density fat fraction (PDFF, %) [25]. This sequence differs from the classical 2‐point Dixon scan by including 3 image sequences (so‐called 3‐point Dixon), which enables more accurate fat fraction (FatFrac) estimation by correcting for magnetic field inhomogeneities and improving image uniformity and PDFF reliability [26]. Patients with evidence of pancreatic diseases (mass lesion, acute or chronic pancreatitis, prior pancreatic surgery, pancreatic duct stenting, or peripancreatic fluid collections) reported in the MRI report were excluded. Using 60 randomly selected MRI scans from this cohort, the pancreas was annotated by expert radiologists (H.T. and N.T.). All slices that included the pancreas were segmented using the open‐source segmentation tool ITK‐Snap [27]. Annotations were performed on fat suppression T1‐weighted post‐contrast images. Initial pilot work had shown that direct segmentation on the IDEAL sequence images yielded poor human‐model segmentation agreement due to the relatively poor resolution of the IDEAL sequences, hence the need for initial T1 post‐contrast segmentation [28]. The nnU‐Net framework (convolutional neural network, CNN) was employed to train a deep learning model for automated pancreas segmentation [29]. The model was trained on 30 annotated images using five‐fold cross validation. Model segmentation performance was evaluated using the Dice similarity coefficient (DSC) on a held‐out test set of the remaining 30 scans. Next, the T1‐weighted MRI‐derived masks were spatially aligned to IDEAL sequence images for each individual patient. The masks were resampled to match the spatial resolution and geometry of the IDEAL images. To ensure spatial correspondence between the T1 mask and the IDEAL image, the mask was reformatted using the Resample function from the SimpleITK library. The IDEAL image was used as the fixed reference space, defining the output grid's origin, spacing, size, and orientation. The T1 mask was resampled onto this grid using nearest‐neighbor interpolation (sitk.sitkNearestNeighbor). The registered pancreatic binary mask was applied to the IDEAL PDFF map (element‐wise multiplication) to retain pancreatic voxels only. IPF was quantified as the mean PDFF within the masked region based on the IDEAL FatFrac sequence. The mean whole‐pancreas PDFF values were calculated by combining the PDFF values from all axial pancreas slices in each individual patient and then compared to corresponding whole‐pancreas PDFF values derived from the human‐annotated images using intraclass correlation coefficient (ICC) and mean absolute error (MAE). Agreement was evaluated across three raters: two expert radiologists and the AI‐generated segmentation, using per‐subject mean PDFF values from 30 studies. A two‐way random‐effects model with absolute agreement was used based on single measurements. Supporting Information S1: Figure S1 shows examples of pancreas segmentation on T1 post‐contrast MRI images by radiology experts and the AI model in the test set. Figure 2 presents an example of an MRI IDEAL Fat Fraction sequence for a segmented pancreas. MRI IDEAL Fat Fraction Image Visualization for Pancreas Segmentation and IPF estimation demonstrating heterogenous distribution of IPF across the different axial slices in the same patient: (a) 3D MRI Voxels for Each Slice (b) Mean IPF % per Slice. This demonstrates the necessity of a whole‐organ quantification approach over an ROI‐based one. IDEAL = iterative decomposition of water and fat with echo asymmetry and least squares estimation; IPF = intrapancreatic fat; ROI = region of interest; PDFF = proton density fat fraction. Step 2: Comparing IPF Quantity and Distribution in HRIs and Non‐HRI Controls IPF Quantification Analysis High risk individuals over the age of 18 were identified from the Mayo Clinic Rochester Prospective Registry for Patients at High Risk for Pancreatic Cancer (HRPC Registry IRB# 20‐012752). The registry includes HRIs with a family history of PC and/or presence of pathogenic germline variants (PGV). Specific inclusion criteria are outlined in Appendix A Supporting Information S1. Patients with PC, acute pancreatitis, endoscopic stenting and altered anatomy secondary to prior pancreatic surgery were excluded. HRIs were matched with non‐HRI controls, selected from the IDEAL cohort in step 1, in a 1:2 ratio matched to HRIs for variables known to affect IPF, including age, sex, body mass index (BMI), and presence of diabetes mellitus (DM). The non‐HRI controls were referred to imaging for reasons other than pancreatic diseases. HRI cases and non‐HRI controls were non‐overlapping for these 2 sequential steps of IPF model development and IPF comparison between cases and controls. Statistical Analysis Patient characteristics were summarized using descriptive statistics. Continuous variables, including age and BMI, were reported as medians with interquartile ranges, means with standard deviations (SD), and ranges. Categorical values such as DM status and sex were presented as frequencies and percentages. Associations between whole‐pancreatic PDFF and clinical variables were evaluated using linear regression models. To evaluate the association of PDFF with high‐risk status, a linear regression model adjusted for age, BMI, DM and sex was fitted. Another similar regression model was used to evaluate the association of PDFF and type of risk (familial vs. genetic). Statistical analyses were performed in R version 4.4.1 (R Core Team; R Foundation for Statistical Computing, Vienna, Austria). IPF Distribution Analysis There is evidence that several IPF spatial distribution patterns exist and that these may have clinical significance [22]. Clustering or comparison of IPF distribution has not been previously reported in the literature; hence, we developed and adopted a novel method. First, each pancreatic segmentation mask and its corresponding IPF distribution were visually inspected (T.K). Cases with fewer than five axial slices (such as cases with very large pancreatic cysts) were excluded (n = 29), as these could distort the apparent distribution pattern and disproportionately influence the analysis. This was done in order to capture the most informative part of the gland while giving every patient an equally dimensioned input, thereby reducing inter‐patient variability and standardizing the clustering input. Segmented FatFrac IDEAL sequences were used to analyze fat distribution. Pixel‐level maximum and mean intensity values were extracted to generate two separate 2D images per subject (examples shown in Supporting Information S1: Figure S4), which were then normalized and centered prior to further analysis. An institutionally approved foundational model, Google Gemini Embeddings for Multimodal (multimodalembedding@001), was employed to extract embedding vectors for each of the normalized images [30]. These embeddings were subsequently analyzed using principal component analysis (PCA) for visualization [31]. Additionally, Fuzzy C‐means clustering was applied to the embedding vectors to identify potential groupings within the data [32, 33]. This approach allowed us to compare the overall pattern of fat distribution across patients and identify patients with similar distribution profiles. Surface Erosion of 3D Pancreas Segmentation Masks Retroperitoneal fat surrounds the pancreas and can impede the accuracy of IPF measurement when using AI‐based pancreas segmentation approaches as partial inclusion of peripancreatic fat may overestimate IPF PDFF. To address this, we performed a boundary erosion analysis. This approach was designed to reduce the contribution of erroneously included peripancreatic fat while preserving true IPF within the central pancreatic parenchyma. Two radiologists (R1 and R2) manually segmented the pancreas on test set MRIs (n = 30) to create human‐derived three‐dimensional (3D) pancreas masks. Three erosion approaches were applied to the outermost layers of the corresponding AI‐generated 3D pancreatic masks that had previously been generated for this test set: (1) uniform one layer‐voxel erosion, (2) one‐layer and (3) two‐layer voxel erosion selectively excluding voxels with PDFF > 50%. Mean absolute error (MAE) in PDFF (measured as %) was compared across these three approaches with the human—derived IPF measurement as ground truth. Results AI Model for IPF Quantification Each pancreatic slice displays a distinct average PDFF (see Figure 2), reaffirming that assessing only specific regions of the pancreas may not yield accurate whole‐organ PDFF measurement. The pairwise DSC of the model on the test set MRI images, calculated versus each of the expert annotators, on the test set, ranged between 70.4% and 89.9% with a mean of 78% (SD 1.8) (Supporting Information S1: Figure S2). Whole‐pancreas PDFF measurements derived from AI‐generated pancreas masks demonstrated excellent concordance with manual annotations by two expert radiologists (Supporting Information S1: Figure S3), with a two‐way ICC of 0.99 (95% CI: 0.981–0.995). The MAE in whole‐pancreas PDFF estimation was 1.4% for both expert radiologists when compared with the AI model (Supporting Information S1: Figure S2). Computational requirements and costs are described in Appendix B in the Supporting Information S1. Comparing IPF Between HRIs and Non‐HRI Controls The study included 255 patients, comprising 85 HRIs and 170 matched non‐HRI controls. 76.5% of the HRI patient scans were performed using the 3.0 T compared with 41.8% of the matched non‐HRI patient scans (the full list of technical details is found in Supporting Information S1: Tables S2). A manual review of all the segmentation masks and registrations revealed errors in 1 HRI case and 4 controls. A post hoc analysis excluding these cases and controls did not significantly alter any study result (Supporting Information S1: Tables S3 and S4). To further assess the robustness of PDFF estimation, we evaluated 3 cases in which model‐based pancreas segmentation was identified as suboptimal on manual review. These cases were independently segmented by two expert radiologists. Despite lower expert–AI DSC in these cases (0.30 ± 0.19 and 0.30 ± 0.20 for experts A and B, respectively), the MAE between model‐derived and expert‐derived whole‐organ PDFF remained low (2.22 ± 0.77 and 2.38 ± 1.26 for experts A and B, respectively), consistent with the original test‐set analysis (Supporting Information S1: Table S5). The median (25th, 75th percentiles) age was 61 (52, 67) years and the median BMI was 27.4 (24.3, 32.5) kg/m2; 52% of the patient population were women and 12.9% had DM (shown in Table 1). TABLE 1Baseline characteristics of HRIs and matched non‐HRI controls.HRIs (N = 85)Controls (N = 170)Total (N = 255)AgeMedian61.061.061.0Q1, Q3 (IQR)52.0, 67.0 (15.0)52.2, 67.0 (14.7)52.0, 67.0Range31.0–75.040.0–75.031.0–75.0SexFemale45 (52.9%)90 (52.9%)135 (52.9%)Male40 (47.1%)80 (47.1%)120 (47.1%)BMI (kg/m2)Median27.627.327.4Q1, Q3 (IQR)24.4, 32.5 (8.1)24.3, 32.4 (8.1)24.3, 32.5 (8.2)Range18.9–52.918.5–44.318.5–52.9Diabetes mellitusAbsent74 (87.1%)148 (87.1%)222 (87.1%)Present11 (12.9%)22 (12.9%)33 (12.9%)Abbreviations: BMI = body mass index; HRIs = high‐risk individuals. Baseline characteristics of HRIs and matched non‐HRI controls. Abbreviations: BMI = body mass index; HRIs = high‐risk individuals. The mean whole‐pancreas PDFF was 25.8% (± 11.5%) and 26.9% (± 11.4%) for HRIs and non‐HRI controls, respectively (Table 2 and Figure 3), without any significant difference between the groups (p = 0.32). Among HRIs, Pancreas PDFF was similar between individuals with PDAC family history and germline variants (p = 0.78). PDFF was significantly associated with age (p = 0.015), male sex (p < 0.001), DM presence (p = 0.009), and BMI (p < 0.001), as shown in Figure 4. TABLE 2Whole‐organ PDFF estimation in HRIs and non‐HRI matched controls.Whole pancreas PDFFHRIs (N = 85)Controls (N = 170)Total (N = 255) p‐valueMean (%, SD)25.8 (11.5)26.9 (11.4)26.6 (11.4)0.32Median (%)25.925.725.8Q1, Q3 (IQR)16.6, 35.1 (18.4)18.4, 33.5 (15.1)18.2, 33.8 (15.5)Range5.0–54.55.0–72.95.0–72.9 Note: Conditional Logistic Regression, conditioned on case‐control set and adjusted for age, sex, BMI and DM.Abbreviations: BMI (kg/m2) = body mass index; DM = diabetes mellitus; HRIs = high‐risk individuals; PDFF = proton density fat fraction. Whole‐organ PDFF estimation in HRIs and non‐HRI matched controls. Note: Conditional Logistic Regression, conditioned on case‐control set and adjusted for age, sex, BMI and DM. Abbreviations: BMI (kg/m2) = body mass index; DM = diabetes mellitus; HRIs = high‐risk individuals; PDFF = proton density fat fraction. Box plot comparing whole‐pancreas Proton Density Fat Fraction (PDFF, in %) between high‐risk individuals (HRIs) and non‐HRI controls. Solid line indicates the median Pancreatic PDFF, and the boxes represent the interquartile range (25th–75th percentile). Whiskers extend to 1.5 times IQR. Dots correspond to whole‐organ mean pancreatic PDFF for individual patients. Correlation between whole pancreatic PDFF and age, BMI, sex and DM status. Scatter plot showing the relationship between PDFF (%) and (A) age (years), (B) BMI (kg/m2). Box and whisker plots show the association between whole pancreatic PDFF and (C) DM and (D) Sex. High‐risk individuals (HRIs) and non‐HRIs represented in red and green, respectively. BMI = body mass index; DM = diabetes mellitus; PDFF = proton density fat fraction. Surface Erosion of 3D Pancreas Segmentation Masks The mean (SD) Pancreas‐PDFF was 22.35 (12.57), 22.20 (11.96) and 23.06 (12.37) for R1, R2 and the AI model, respectively. Uniform one‐layer erosion did not significantly lower MAE for IPF measurement. However, both one‐layer and two‐layer selective erosion approaches removing voxels with PDFF > 50% significantly reduced the MAE for human‐AI comparisons. The MAE (SD) in PDFF for R1 versus AI and R2 versus AI was 1.38 (1.15) and 1.42 (1.37) without erosion; 1.21 (1.02) and 1.62 (1.18) with uniform one‐layer erosion (p = 0.48 and 0.27); 0.92 (0.79) and 1.15 (1.14) with one‐layer selective erosion (p = 0.01 for both) and the lowest for the two‐layer selective erosion approach: 0.79 (0.74) and 0.88 (0.88) (p < 0.001 for both) (Supporting Information S1: Table S6). In post hoc analyses using the corrected PDFF values, median PDFF was lower in both groups but remained broadly similar between HRIs and matched non‐HRI controls (Supporting Information S1: Table S7 and Figure S7). None of the other results (PDFF vs. Age, BMI, Sex, DM, Type of HRI) changed (Supporting Information S1: Table S8). However, voxel counts differed between groups, with lower voxel counts among HRIs than non‐HRIs, both before and after erosion (Supporting Information S1: Figure S8 and Table S9). Exploratory assessment of the relationship between voxel count and PDFF suggested that, above approximately 2500 voxels, the association between voxel count and PDFF was similar in HRIs and non‐HRIs (Supporting Information S1: Figure S9). Below this range, PDFF appeared to decrease as voxel count decreased among HRIs, a pattern that was not clearly observed among non‐HRIs. However, the number of HRIs in this low‐voxel range was small, precluding further inferential analysis and these findings should therefore be interpreted with caution. PCA visualization of the image embeddings did not reveal any distinct clustering or separation between HRIs and non‐HRIs, suggesting that the model did not identify overt systematic differences in spatial fat distribution patterns between the groups. Additionally, no distinct clustering patterns were observed when the data were stratified by age, sex, presence of DM, or BMI (Supporting Information S1: Figures S5 and S6). These findings suggest no clear difference in the spatial distribution patterns of IPF between the two groups. Discussion This two‐step study describes the development of an AI‐based tool for whole‐organ quantification of IPF on MRI, incorporating both quantitative measurement and spatial distribution analysis. Using this tool in a matched case‐control design, we found that individuals at high‐risk for PC due to familial and/or germline predisposition had comparable IPF quantity and distribution to matched non‐HRI controls. To our knowledge, this is the first study to systematically evaluate whole‐organ IPF quantity and spatial distribution in individuals with familial or germline risk of PC. Although MRI is the preferred imaging modality for IPF quantification, correlating with histology [8], published methods for IPF quantification vary widely [1, 4, 7, 11, 34]. These remain limited by ROI‐based or slice‐based measurements that are vulnerable to sampling bias because IPF is heterogeneously distributed throughout the pancreas [9, 10, 12]. These limitations have hindered the development of reproducible IPF assessment in clinical practice and large‐scale research. In this context, automated whole‐organ quantification represents an important advance, enabling objective, scalable, and more representative estimation of pancreatic fat [9, 14]. Previous automated MRI approaches for IPF assessment have advanced the field but remain limited by reliance on 2‐point Dixon imaging, regional or slice‐based quantification, non‐dedicated fat sequences, or single‐field‐strength datasets [35, 36, 37, 38]. In contrast, our approach combines automated whole‐pancreas segmentation with IDEAL‐based PDFF quantification, enabling objective whole‐organ IPF assessment using a clinically acquired quantitative fat sequence [39]. Direct anatomical segmentation on IDEAL images is limited by poor soft tissue contrast, which can reduce inter‐observer agreement and segmentation reliability [9, 40, 41]. To overcome this limitation, we used a two‐step approach in which segmentation was first performed on fat‐suppressed post‐contrast T1‐weighted images, which provide clearer pancreatic borders and are routinely acquired in contrast‐enhanced abdominal MRI [28]. The resulting masks were then registered in the IDEAL MRI space, allowing anatomical segmentation from the T1‐weighted sequence to be transferred onto the quantitative fat‐water imaging domain. This approach combines the superior anatomical resolution of post‐contrast T1‐weighted imaging with the quantitative fat‐fraction capabilities of IDEAL, enabling more reproducible whole‐organ IPF estimation. Additionally, human expert verification of accurate mask registration added internal validity to the approach in this study. The finding that HRIs and non‐HRI controls have comparable quantities of IPF when matched for age, sex and BMI advances our knowledge of the interaction between IPF and known PC risk factors. This is relevant as emerging evidence suggests that IPF may play a causative role in pancreatic carcinogenesis [21]. A recent Mendelian randomization study based on large scale genome‐wide association studies for MRI‐measured IPF using the UK Biobank suggested a causative relationship. Using 8 single nucleotide polymorphisms (SNPs) that strongly correlated with IPF, each +1 SD (∼7.9%) genetic increase in IPF was linked to a 52% higher PC risk (OR 1.52, 95% CI 1.15–2.00; p = 0.003) [13]. Supporting this, animal and human histopathologic studies have linked pancreatic fatty infiltration with precursor neoplastic changes, including TGFβ‐mediated acinar‐to‐ductal metaplasia, facilitation of KRAS‐induced neoplasia, and increased PanIN burden [23, 42]. Our initial approach of using the entire segmentation mask for IPF quantification resulted in mean PDFF numbers that were higher in both groups compared to what has been traditionally reported in the literature, although the absence of a clear consensus on normative PDFF thresholds limits the interpretation of absolute IPF values. While it is possible that this may be related to the high median BMI across both groups in our study population, we considered the possibility that partial inclusion of peripancreatic fat may overestimate PDFF in this segmentation‐based approach for IPF quantification. To address this, we performed a surface voxel erosion analysis with the goal of minimizing the inclusion of peripancreatic fat while preserving the representation of pancreatic parenchyma. Previous studies have addressed this potential confounder using either complete erosion of the outermost layer of the 3D mask or exclusion of all voxels with PDFF > 50% throughout the mask [3, 35]. However, these arbitrary approaches may remove voxels containing true IPF. We therefore applied a selective threshold‐based approach, excluding voxels with PDFF > 50% only within the two outermost mask layers. Unlike previous studies, in which methods were selected based on arbitrary investigator decisions, we compared several approaches and selected the one that most improved agreement between automated IPF estimates and measurements by two independent expert radiologists. In this post hoc analysis after voxel erosion HRIs had lower mean PDFF compared with non‐HRI matched controls. This difference appeared to be driven by total voxel count, indicating that HRIs with lower pancreatic voxel counts may have less fat compared to non HRIs with a similar number of voxels. However, the number of HRIs in this low‐voxel range was small, which prevented further inferential analysis. We conclude that this study shows that IPF quantity is similar among groups with and without genetic and/or familial predisposition to PC, although the potential influence of voxel count, possibly as a surrogate for pancreatic volume, warrants further investigation. Limited evidence suggests that IPF may follow distinct spatial distribution patterns within the pancreas, although prior regional studies have reported inconsistent findings regarding the predominant location and clinical correlates of fat deposition [22, 43, 44]. These discrepancies may reflect the limitations of sampling‐based approaches, which may not fully capture heterogeneous whole‐organ fat distribution. Therefore, in addition to quantifying total IPF, we applied a novel whole‐organ embedding‐based approach to evaluate spatial distribution patterns. We found no distinct separation between HRIs and non‐HRI controls, nor clear clustering by age, sex, DM, or BMI. Although exploratory and not specifically powered for these secondary analyses, these findings suggest that IPF spatial distribution does not differ meaningfully by familial or germline PC risk status and should be further evaluated in larger population‐based cohorts. Our study has limitations. First, the AI model was trained (n = 30) and tested (n = 30) on a relatively small number of scans, which may limit generalizability. However, segmentation performance was comparable to prior neural network approaches, where DSC for CNNs on larger MRI datasets ranged between 0.69 and 0.80 [45, 46] and more importantly, the downstream quantification model was highly accurate with less than 2% error for IPF quantification in the test set. Second, our two‐step approach relies on the accurate co‐registration of T1‐weighted and IDEAL sequences. While this is a necessary step, as with other models, minor registration errors could potentially introduce inaccuracies in the final PDFF map as well as potential biases related to scanner field strength and acquisition heterogeneity. Third, this was a single‐center study, although cases and controls were rigorously matched for key IPF‐associated factors, and scans were acquired across different MRI platforms and field strengths. Ethnic variation, which might also influence IPF, was not evaluated and will need to be carefully addressed in future population‐based cohort studies on intrapancreatic fat. Finally, the spatial distribution analysis used a novel embedding‐based approach that remains exploratory and requires validation in larger cohorts. In conclusion, using a novel and accurate AI‐based tool, we provide a comprehensive analysis of whole‐organ IPF quantification and distribution patterns in HRIs and matched non‐HRI controls. Our early findings show no substantial differences in either IPF quantity or spatial distribution between these groups; however, larger studies are needed to validate these findings, further refine the approach for AI‐based IPF measurement and establish population based diagnostic thresholds for the diagnosis of fatty pancreas. The results of our study generate the hypothesis that IPF‐driven carcinogenesis and germline‐driven carcinogenesis are possibly two independent pathways to pancreatic cancer and highlight the need for further investigation into the metabolic determinants of IPF and its role in carcinogenesis. Importantly, the AI‐based tool developed here provides a scalable and objective method for whole‐gland IPF quantification, enabling large‐scale population studies that were previously infeasible and present a potential pathway for clinical implementation of IPF measurement in routine abdominal MRI. Future work should focus on applying this tool to longitudinal cohorts and different populations to understand how changes in IPF over time relate to metabolic health and the risk of PC. Author Contributions Tom Konikoff: conceptualization (supportive), data curation (equal), formal analysis (supportive), investigation (equal), methodology (equal), visualization (lead), writing – original draft (lead). Laurens P. Janssens: conceptualization (equal), data curation (equal), formal analysis (supportive), investigation (equal), methodology (equal), visualization (equal), writing – original draft (supportive), writing – review and editing (equal). Hiroaki Takahashi: conceptualization (equal), data curation (equal), methodology (equal), writing – review and editing (equal). Garima Suman: data curation (equal), methodology (supportive), writing – review and editing (supportive). Fred Nugen: methodology (supportive), software (supportive), writing – review and editing (equal). Anup Mishra: conceptualization (equal), formal analysis (equal), methodology (equal), software (lead), visualization (lead), writing – original draft (equal). Panos Korfiatis: methodology (equal), software (supportive), writing – review and editing (equal). Erin E. Carlson: formal analysis (equal), writing – review and editing (equal). William B. Bamlet: conceptualization (equal), formal analysis (equal), writing – review and editing (equal). Aishwarya Pradeep: data curation (equal), writing – review and editing (equal). Kelly Harper: data curation (supportive), methodology (supportive), writing – review and editing (equal). Karen A. Doering: project administration (equal), writing – review and editing (supportive). Adriana Delgado: project administration (lead), writing – review and editing (supportive). Bradley J. Erickson: conceptualization (equal), methodology (equal), supervision (equal), writing – review and editing (equal). Ajit H. Goenka: conceptualization (equal), writing – review and editing (equal). Naoki Takahashi: conceptualization (equal), data curation (equal), investigation (equal), methodology (equal), supervision (equal), writing – review and editing (equal). Ann L. Oberg: conceptualization (equal), formal analysis (lead), investigation (equal), methodology (equal), supervision (equal), writing – review and editing (equal). Shounak Majumder: conceptualization (lead), funding acquisition (lead), investigation (equal), methodology (equal), supervision (lead), writing – review and editing (lead). Funding Funding was provided by the Centene Foundation (SM). S.M. is also supported by U01 CA210138, National Cancer Institute. Ethics Statement Access to deidentified datasets may be made available upon reasonable requests to the corresponding author, subject to Mayo Clinic institutional review board review and approval. Conflicts of Interest The Mayo Clinic and Exact Sciences have an intellectual property development agreement. S.M. is listed as inventor under this agreement and could share potential future royalties as an employee of Mayo Clinic. A.H.G. is supported by National Institutes of Health (Grant Numbers R01CA272628 and R01CA256969), Mayo Clinic Comprehensive Cancer Center, Champions for Hope Pancreatic Cancer Research Program of the Funk Zitiello Foundation, David E. Simon & Jacqueline S. Simon Charitable Foundation, Centene Foundation, and Hoveida Family Foundation. Consultant for Ferronova and GE Healthcare. DSMB Member, Early Detection Initiative (EDI), PanCan. All other authors declare no conflicts of interest. Supporting information Supporting Information S1 Data Availability Statement The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions. References Associated Data Supplementary Materials Supporting Information S1 Data Availability Statement The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.