Introduction

Intraductal papillary mucinous neoplasms (IPMNs) of the pancreas are the most frequently identified precursor lesions of pancreatic cancer (PDAC). Only a minority progress towards malignancy, making their identification a critical point in cancer prevention. The available guidelines are imprecise, commonly resulting in over treatment of indolent lesions.

Aims & Methods

The aim of this study (EMI-IPMN, NCT06706700) is to discover novel biological and radiological biomarkers that can distinguish malignant from benign IPMNs. We focused on Endocrine, Metabolic and Inflammatory (EMI) factors, some pathways that are characterized in PDAC but poorly investigated in IPMN progression.
Patients with a presumed diagnosis of IPMN with indication for surgical resection who underwent high-quality magnetic resonance imaging (MRI) and/or computed tomography (CT), were prospectively enrolled. Peripheral blood (PB) and cyst fluid (CF) samples were collected during surgery, and lesions were histologically classified as IPMN with low-grade dysplasia (LGD), high-grade dysplasia (HGD), or invasive cancer (IC). Malignant IPMNs were defined as those with HGD and IC. Cytokine and hormone levels were measured in both fluids using multiplex assays and compared across the diagnostic groups. Continuous variables were converted into categorical variables based on optimal cut-off values, and both univariate and multivariate logistic regression analyses were conducted.

Results

Among the 205 patients enrolled, 133 (58 females, 75 males, median age 69.2) had a histological diagnosis of IPMN: 49 (36.8%) had LGD, 44 (33.1%) HGD and 40 (30.1%) had IC. Analysis of the EMI profile revealed higher concentrations of several markers (IL-8, VEGF, GLP-1, glucagon, and C-peptide) in CF compared to PB. Specific cytokines such as MCP-1 IP-10, TNFα, and IL-8 were significantly elevated in the CF of HGD/IC cases compared to LGD (p=0.05, 0.041, 0.001 and 0.025 respectively). Similarly, IP-10, IL-8, and MDC levels in PB were higher in HGD/IC patients (p=0.006, 0.008 and 0.004 respectively).
Pancreatic fat content, assessed with mDixon Quant Fat Fraction MRI sequences, was notably higher in HGD/IC patients (13.89%) compared to LGD patients (8.38%, p<0.05), and positively correlated with MCP-1 levels in CF.
Among these biological and radiological markers TNFα in the CF demonstrated a substantial diagnostic accuracy for identifying malignant IPMN, with an area under the ROC curve (AUC) of 0.72. In the final multivariable analysis we demonstrate that high TNFα levels in CF and HbA1c were the two independent risk factors for the identification of malignant IPMN (OR 10.8 and 1.11 respectively).

Conclusion

The progression of IPMN to PDAC is associated with a proinflammatory shift in the tumor microenvironment, reflected by local (CF) and systemic (PB) elevations in key cytokines (TNFα, MCP-1, and IL-8) whose involvement in PDAC progression is well established. Elevated hormone levels in CF, along with the strong association between malignant IPMN and HbA1c, underscore the critical role of endocrine alterations in the early stages of pancreatic tumorigenesis. These findings may be harnessed to enable earlier detection of malignant transformation in IPMN. Moreover, the correlation between imaging markers and biological profiles suggests that MRI could serve as non-invasive proxies for underlying pathological changes. Together, these results offer new insights into the pathophysiology of IPMN progression and may imporve diagnostic strategies

Introduction

Early diagnosis of pancreatic cancer remains difficult due to the lack of overt findings on imaging. Our prior analysis revealed that indirect signs such as pancreatic atrophy and ductal stenosis or dilation often precede tumor visibility. In collaboration with Fujifilm Corporation, we developed an AI-based system capable of detecting pancreatic masses, atrophy, and ductal abnormalities on contrast-enhanced CT(CE-CT) and non-contrast CT(NC-CT). This study evaluated its diagnostic performance.

Aims & Methods

The AI algorithm extracts the pancreatic parenchyma, ducts, and tumor regions from CT images and classifies the presence of four imaging findings (pancreatic mass, atrophy, ductal dilation, and ductal stenosis). The model was trained and validated using approximately 2,000 images. For testing, we used an external dataset comprising 90 pancreatic cancer cases (Tis/T1: T2: T3–T4 = 1:1:1) and 100 controls collected from affiliated institutions. Reference standard for each imaging findings was established by two expert radiologists. Diagnosis of pancreas cancer was also performed using the logistic regression model with the results of each imaging finding. Diagnostic performances were performed on CE-CT and NC-CT, and AI diagnostic accuracy and reading time were compared with those of four radiologists and two gastroenterologists.

Results

For the diagnosis of pancreas cancer, the AI showed a sensitivity of 98.9% and specificity of 93.0% on CE-CT, and sensitivity of 83.9% and specificity of 92.0% on NC-CT. The system successfully detected early-stage cancers, including Tis and T1, by identifying indirect imaging findings. Detection metrics for CE-CT vs. NC-CT were as follows: mass (sensitivity 97.2% vs. 87.5%; specificity 74.6% vs. 71.4%), atrophy (74.0% vs. 59.6%; 90.7% vs. 92.9%), ductal dilation (78.8% vs. 90.5%; 96.4% vs. 88.3%), and ductal stenosis (87.7% vs. 93.9%; 90.4% vs. 85.1%).
Physicians showed mean sensitivities/specificities of 88.9%/99.3% on CE-CT and 53.4%/99.5% on NC-CT. Mean reading time per case was 22 seconds (AI) vs. 63 seconds (physicians) for CE-CT and 20 seconds (AI) vs. 78 seconds (physicians) for NC-CT.

Conclusion

The AI system demonstrated diagnostic performance comparable to that of experienced radiologists and significantly reduced reading time. While its specificity was somewhat lower than that of physicians, the system effectively identified early-stage pancreatic cancers using secondary findings. At present, it may be best applied as a second-reader support tool in radiological interpretation. Further refinement may expand its role in early detection and possibly future screening.

Disclosure

N.H., A.O., and M.T. are employees of Fujifilm Corporation.
A.M., K.S., T.M., and Y.K. declare financial conflicts of interest with Fujifilm Corporation.

Introduction

Any computer-assisted characterization (CADx) system relies on Artificial Intelligence algorithms to provide a probability score to the endoscopist, characterizing the polyp as an adenoma or a hyperplastic polyp.^1-3 However, the clinical utility of these AI-generated outputs is dependent on the endoscopist’s ability to interpret and integrate them into their final diagnostic decision-making process. This requires not only recognizing and accepting a true positive AI diagnosis, but also accurately identifying and rejecting false positives. This requires a higher level of endoscopist expertise to discern whether an AI-generated classification is accurate and should be incorporated into the final diagnosis or requires correction. The primary challenge in CADx implementation lies in this inherent complexity of human-AI interactions, which can lead to misinterpretation, resulting in the endoscopist either rejecting correct CADx outputs or accepting incorrect classifications, compromising diagnostic accuracy.

Aims & Methods

We conducted a secondary analysis from a prospective study. Endoscopists were first presented with the CADx classification for each polyp, which they could accept or reject and provide an alternative. A total of 1143 patients aged at least 45 underwent colonoscopy between June 2022 and January 2025 at the CHUM Hospital, yielding 753 polyps from 367 patients after exclusion criteria. The primary aim was to assess the agreement/rejection patterns of the CADx diagnosis by the endoscopists, stratified by categorization. Secondary outcomes included the rate at which clinicians correctly rejected inaccurate CADx outputs, as well as trends in post-rejection reclassification.

Results

A total of 89.1% (408/458, 95% CI: (85.8,91.7)) of accepted CADx neoplastic diagnoses were correct, and 68.7% (68/99, 95% CI: (58.5,77.4)) of accepted CADx hyperplastic diagnoses were accurate. However, endoscopists accepted incorrect diagnoses significantly more often when CADx predicted hyperplastic than neoplastic (%, 95% CI: 31.3% (22.6,41.5) vs 10.9% (8.3,14.2)). Regarding the post-rejection patterns, when endoscopists rejected adenoma classifications (N=47), they falsely classified 25 cases (53.2%, 95% CI: (38.2,67.6)) as a sessile serrated lesion/polyp (SSL/P). Similarly, although 43.6% of the rejected hyperplastic CADx predictions were correctly re-assigned an adenoma (31.5%, 95% CI: (24.3,39.7)) or SSL/P diagnosis (12.1%, 95% CI: (7.5,18.7)), there were 41.6% (95% CI: (33.7,50.0)) that were incorrectly re-classified as SSL/Ps.

Conclusion

Endoscopists generally accept accurate adenoma predictions provided by CADx systems but frequently misclassify hyperplastic polyps as SSLs, demonstrating a diagnostic bias toward neoplasia. Although beneficial from a safety perspective, this bias may result in unnecessary early follow-up colonoscopies, increasing associated costs.^4,5 Developing and integrating CADx systems with robust SSL recognition capabilities seems essential to overcome these diagnostic challenges and enhance clinical decision-making when using CADx assisted optical polyp diagnosis. Such advanced CADx systems are likely crucial for successful CADx integration into routine clinical practice.

References

1. Djinbachian R, Taghiakbari M, Barkun A, et al. Optimized computer-assisted technique for increasing adenoma detection during colonoscopy: a randomized controlled trial. Surg Endosc 2025;39:1120-1127.
2. Byrne MF, Von Renteln D, Barkun AN. Artificial Intelligence-Aided Colonoscopy for Characterizing and Detecting Colorectal Polyps: Required, Nice to Have, or Overhyped? Gastroenterology 2023;164:332-333.
3. Taghiakbari M, Mori Y, von Renteln D. Artificial intelligence-assisted colonoscopy: A review of current state of practice and research. World J Gastroenterol 2021;27:8103-8122.à
4. Force AAT, Parasa S, Berzin T, et al. Consensus statements on the current landscape of artificial intelligence applications in endoscopy, addressing roadblocks, and advancing artificial intelligence in gastroenterology. Gastrointest Endosc 2025;101:2-9 e1.
5. Halvorsen N, Barua I, Kudo SE, et al. Leaving colorectal polyps in situ with endocytoscopy assisted by computer-aided diagnosis: a cost-effectiveness study. Endoscopy 2025.

Disclosure

Daniel von Renteln has received research funding from ERBE Elektromedizin GmbH, Ventage, Pendopharm, Fujifilm and Pentax, and has received consultant or speaker fees from Boston Scientific Inc., ERBE Elektromedizin GmbH, and Pendopharm. Roupen Djinbachian has received speaker fees from Fujifilm. The remaining authors declare that they have no conflict of interest.

Introduction

Computer-aided diagnosis (CAD) systems are widely implemented in the medical field for the analysis of medical images. PolyDeep is a CAD system that has demonstrated its efficacy in lesion detection.

Aims & Methods

The aim of this study is to determine whether PolyDeep increases the adenoma detection rate (ADR) in the context of expert endoscopists participating in a regional, population-based colorectal cancer screening program in Spain. PolyDeep Advance 3 (NCT05513261) is a multicentric, randomized controlled clinical trial. For the primary endpoint, we aimed to assess whether the ADR is superior in PolyDeep-assisted colonoscopy compared to conventional colonoscopy. As secondary endpoints, we aimed to determine whether PolyDeep-assisted colonoscopy improves the polyp detection rate (PDR), serrated lesion detection rate (SLDR), advanced serrated lesion detection rate (ASLDR), and advanced adenoma detection rate (AADR). We included participants who underwent colonoscopy for screening (i.e. those with a positive faecal immunochemical occult blood test – FIT+) or for surveillance after resection of advanced adenomas. We compared both groups using the chi-square and t-Student test.

Results

Between November 2023 and May 2024, we recruited 857 patients, of whom 30 were excluded. The baseline characteristics of the 827 patients included in the analysis were similar (male: 59.3%, screening colonoscopy: 66.7%, mean age: 61.7 ± 6.5 years, withdrawal time: 10:49 ± 04:32 minutes, colorectal cancer incidence: 2.8%). However, there were significant differences between both groups in the Boston Bowel Preparation Scale (conventional colonoscopy: 7.7 ± 1.3 vs. PolyDeep-assisted colonoscopy: 7.4 ± 1.3, p = 0.01). Regarding ADR, we did not find statistically significant differences between conventional and PolyDeep-assisted colonoscopy (57.7% vs. 60.8%, p = 0.4). Similarly, for the secondary endpoints, there were no significant differences in PDR (68.7% vs. 69.6%, p = 0.8), SLDR (32.0% vs. 29.4%, p = 0.5), ASLDR (5.3% vs. 5.1%, p = 1.0), or AADR (24.5% vs. 23.6%, p = 0.8). In Table 1, we described lesion detection rates between both groups by indication (i.e., FIT+ or surveillance).

Conclusion

In the context of expert endoscopists PolyDeep did not increase the ADR.

Introduction

Endoscopic ultrasound (EUS) is the gold standard for evaluating pancreatic lesions, allowing for real-time imaging and biopsy sampling. Nevertheless, the differentiation between solid pancreatic lesions — such as pancreatic ductal adenocarcinoma (PDAC) and pancreatic neuroendocrine tumors (PNETs) — as well as cystic lesions is often troublesome. Specifically, identifying intraductal papillary mucinous neoplasms (IPMNs) with varying degrees of dysplasia is crucial for guiding clinical management. However, the diagnostic accuracy of current methods remains suboptimal. This multicenter study aimed to develop and validate convolutional neural networks (CNNs) to characterize solid and cystic pancreatic lesions and stratify IPMNs into high-grade dysplasia/carcinoma (HGD/C) and low-grade dysplasia (LGD).

Aims & Methods

This multicenter study included EUS images from 12 centers in Spain, Brazil, the United Kingdom and the United States of America. Solid lesion differentiation (PDAC vs. PNET) was achieved using a CNN trained on 49034 images from 95 patients (n=69 for PDAC and n=26 for PNET), with diagnoses confirmed via EUS-guided fine-needle biopsy or surgical specimen analysis. A second CNN was trained on 79395 images from 82 patients (n=64 for mucinous cystic lesions and n=18 for non-mucinous cystic lesions). Finally, a third model used the subset of IPMNs (n=51046 frames, 30 patients) to distinguish between HGD/C (n=5 patients) and LGD (n=25 patients). Histological confirmation was used as the gold standard. Performance metrics included sensitivity, specificity, accuracy, and area under the precision-recall curve (AUC-PR). To comprehensively determine the model's performance, we evaluated multiple thresholds for the confidence score and the IoU, between 0.45 and 0.55 with a step of 0.02 ([0.45, 0.47, 0.50, 0.52, 0.55]). The datasets were split into training (70%), validation (20%), and testing (10%) subsets.

Results

For solid lesions, the model differentiated PDAC from PNETs with an accuracy of 91.8%, with sensitivities of 96.2% and 97.8%, respectively. The AUC-PR was 0.982 for PDAC and 0.977 for PNET. Regarding cystic lesions the model detected mucinous cystic lesions with a sensitivity of 97.5%, an overall accuracy of 89.7% and an AUC-PR of 0.977. When the subset of IPMNs was specifically analyzed, the model distinguished IPMNs with HGD/C from those with LGD with a sensitivity of 97.5%, specificity of 94.3%, and an overall accuracy of 89.8%. The AUC was 0.988.

Conclusion

This is the first multicenter, transatlantic AI study aiming to provide a holistic approach to both solid and cystic pancreatic lesions, focusing on those with a significantly higher malignant potential. Our models achieved high accuracy in identifying and differentiating these conditions. Furthermore, to our knowledge this is the first integrative model to include the ability to stratify IPMNs into LGD and HGD/C, which has significant clinical implications, potentially reducing unnecessary surgeries and ensuring timely intervention for high-risk patients. The inclusion of data from diverse regions reduces demographic bias and enhances model generalizability. Real-time implementation of these solutions could significantly improve decision-making in EUS and optimize patient outcomes.