Machine Learning-Based Detection of Intestinal Metaplasia Identifies Clinically Missed Cases.
Gastric adenocarcinoma is an aggressive disease with more than 1 million new cancer cases in 2020 and more than 700 000 cancer deaths globally. (1) Gastric intestinal metaplasia (IM) is a relatively common precancerous condition that may progress to dysplasia or gastric carcinoma, though the underlying rate of progression varies by study population. (2) Given the aggressive nature of gastric cancer, targeted disease interception programs for patients with IM is a promising approach to reduce disease risk through endoscopic surveillance and interception of dysplasia and early cancer. (3-7)Gastric IM, which is classically characterized by the presence of goblet cell metaplasia, though potentially also identified by the presence of Paneth cells or brush border-bearing enterocytes, is hypothesized to be a response to chronic injury and is frequently seen in biopsies from patients with concurrent chronic gastritis. Standardized endoscopic collection and pathologic-reporting schemes have been proposed to improve the diagnostic utility of endoscopic gastrointestinal biopsies. (8, 9)
IM is a proposed initial step in the progression of normal gastric epithelium to malignant gastric intestinal-type adenocarcinoma with atrophy and dysplasia. (10, 11) Morphologic subtleties within gastric IM are associated with variation in cancer risk. "Complete-type" IM describes a region of normal gastric epithelium that has been replaced by a continuous length of typical intestinal metaplastic cells, goblet cells, and enterocytes. "Incomplete-type" metaplasia instead demonstrates a heterogeneous mixture of the goblet cells and enterocytes of "complete" metaplasia and gastric epithelium. Incomplete IM is a risk factor for gastric intestinal-type adenocarcinoma, but the 2 types are frequently seen in the same specimens. (11, 12) The location of IM also bears relevance, with extensive IM involving the corpus being associated with the highest risk of progression. (13)
Accurate reporting of IM is essential to effectively characterize IM as a screening tool for gastric carcinoma. Digitization of biopsy slides offers the opportunity to standardize detection and reporting of IM by applying computer vision algorithms to histopathology images. Deep learning-based algorithms have proven particularly adept at this sort of biomedical image processing, with successful applications in diverse pathologic processes. (14-17) In the case of IM, complex algorithms have been implemented. (18, 19) These approaches have occasionally demonstrated detection of previously unknown IM that was missed clinically, but systematic application as a screening step in a quality assurance (QA) pipeline has not been described.
We sought to develop a fully open-source pipeline for the large-scale detection of IM, including minute foci, in clinically acquired biopsy specimens. We further report the pathologic features of clinically missed IM. We explore its utility in real-world clinical use cases, comparing its results to those of practicing gastrointestinal fellowship-trained pathologists and using it to automatically review whole slide images of archival pathology cases. Finally, we demonstrate concurrent interpretation of gastric biopsy diagnostic text to form a proposed QA pipeline.
METHODS
Whole Slide Image Data Set
Two distinct image data sets were used for analysis. The first was comprised of de-identified whole slide images of recut hematoxylineosin-stained (H&E) sections derived from archival tissues from clinical cases (GRACE cohort). Under an institutional review board-approved study, as part of the Duke Gastric Cancer Disparities project (NCIP20, Pro00104555), under a waiver of informed consent, archival tissues from adult patients were accessed as previously described. (20) These slides were scanned on a GT450 (Leica Biosystems) at x40 equivalent resolution (0.26 um per pixel).
Additional slides were identified through the Duke Biorepository and Precision Pathology Center (BRPC) under a waiver of informed consent (Pro00107429) to include patients of all ages with a gastric biopsy specimen at Duke University Hospital (Durham, North Carolina). Clinical databases were searched by the BRPC core staff using custom search tools to identify surgical pathology case parts representing gastric biopsies of any clinical diagnosis. Case final diagnosis text was parsed into specimen headers and specimen diagnoses, and identifying information was stripped. Biopsy specimens with diagnoses of adenocarcinoma, dysplasia, other neoplasia, or gastric polyps were then excluded. Specimen diagnoses were reviewed and validated by a gastrointestinal pathologist.
Specimens were further filtered by those with whole slide images available in a large digital repository generated through an institutional effort to scan millions of slides from the Duke University Health System pathology archives in collaboration with Nference (Cambridge, Massachusetts). Whole slide images were produced by Pramana scanners at x40 equivalent resolution (0.287 um per pixel) and were scanned retrospectively from the archives following prior diagnostic clinical assessment. Some contained ink markings from diagnostic review, which were not removed. Images of slides from gastric biopsy cases were exported with a de-identified ID, using custom scripts and converted from a zero-layer-only DICOM format into TIFF image pyramid format for ease of viewing and annotation.
Both the parsed diagnostic texts and de-identified images from the Nference archival scanning project that were limited to the relevant specimen were provided by using a common de-identified designation.
Active Learning Annotation
To identify specific areas of gastric IM, training was implemented in successive iterations following an "active learning" approach (Figure 1, A and B). Selected GRACE cohort slide images and Nference scans were accrued to a training set and used for annotation. These included a mix of validated images that were both positive and negative for IM, chronic gastritis, and various other nonneoplastic pathologic processes. Annotations of specific areas of IM were performed iteratively with the QuPath viewer. (21) Annotation was either performed by a pathology trainee followed by review by a boarded gastrointestinal specialist pathologist, or directly annotated by the pathologist. A UNet (22) implementation operating on the x4 downsampled layer (equivalent to approximately x10 objective magnification) was trained from random initialization, using a custom PyTorch framework for a variable number of epochs (between 1000 and 200) on progressively expanding subsets of the data. Training and validation sets were randomly split on a patchwise basis for the training steps to monitor model convergence.
A given iteration of the UNet model was then tested with later, unannotated batches of untrained images from the training cohort whole slide images. Annotations resulting from the model were visualized in QuPath as contours. Once visualized in QuPath, annotations were corrected, producing additional training data. Additional data were aggregated with the prior training data for additional training runs, with the previously trained UNet model weights used to initialize the model. Training proceeded iteratively until the entire training batch was annotated (71 total slides).
Diagnostic Model Development
Patches (224 x 224 pixels) derived from the x4 downsampled images of the whole slide image were used with augmentations to include random flips and rotations. Training and validation patches were split patchwise (90:10 split, respectively) to monitor model convergence. Patches were labeled as positive if they contained any pixels overlapping IM annotations. All patches within the annotated areas (including both tissue-containing and non-tissue-containing areas) were used in model training.
Model training implemented transfer learning from ResNet50 (23) with pretrained weights obtained through the torchvision package. (24) The final classifier layer was replaced with an affine linear convolution block with 2 output features. Training was run for 100 epochs on all model weights with observed convergence on the validation set.
Diagnostic Model Implementation
Slide processing began with an Otsu-thresholding-based tissue-finding step using a grayscale downsampled thumbnail image. (25) Non-overlapping patches containing tissue were identified and processed independently through the pretrained model and converted to a scored probability of IM by using the softmax function. Only the initial 2 H&E-stained levels for each block (performed by default on all cases) were provided to the algorithm for review.
Threshold Selection
An initial 150-case cohort (50 IM-positive and 100 IM-negative cases, all distinct from the training data set) was arbitrarily selected from clinical case archives for setting of detection thresholds. The number of "positive" patches exceeding a spread of thresholds (0.5, 0.75, 0.8, 0.9, 0.95, 0.98, 0.99, 0.995, 0.998, and 0.999) and the total number of analyzed (tissue-containing) patches was totaled across all included slides for the part (in some cases including multiple blocks). Area-under-the-curve (AUC) computations were performed at each patch level threshold, examining both raw total patch counts and patch counts normalized to total tissue patches.
Model Testing and Expert Validation in Archival Specimens
Initial H&E levels from the remaining cases not used for training or threshold selection were processed to identify candidate positive patches of IM. Following machine learning analysis, cases were reviewed digitally by 1 pathologist unblinded to the machine learning result (visualized as an IM heatmap overlay of the case) and initial diagnosis. Additionally, all discordant diagnoses, along with 100 arbitrarily selected nondiscordant cases, were reviewed by 3 gastroenterologic pathologists who were initially blinded to both the initial diagnosis and machine learning result. Blinded review additionally included scoring of the abundance of IM in biopsies with positive findings. Subsequently, a final consensus opinion of the status of each case was established by all 4 pathologists reviewing cases unblinded to the initial pathologist impression and machine learning result. Computation of the Fleiss k statistic for interrater concordance of the reviewed specimens was computed in R (version 4.5.1) using the "irr" package.
Large Language Model Analysis of Diagnostic Text
In addition to manual review of the diagnostic text associated with case parts, a large language model (LLM) zero-shot analysis was also attempted with Llama 3.1 8B Instruct. (26) The LLM was prompted with the system role "You are a careful research assistant aiding a pathologist in categorization of gastric biopsies" and the prompt content "Please read the following diagnosis and answer in a single word whether it is positive, negative, or equivocal for intestinal metaplasia:" followed by the diagnosis text. Output was categorized as positive or negative if these precise words were used, and equivocal for any other output. All analysis was performed on institutional machines to prevent disbursement of de-identified patient data.
RESULTS
Model Development
The training set of 71 slide images from gastric IM cases (8 incomplete only, 34 complete only, 9 mixed, and 20 without IM) was annotated with the "active learning approach" (Figure 1, A), yielding 2040 total annotation objects. When applied to the threshold selection cohort of 150 archival specimens (50 with gastric IM and 100 reported as negative), the maximum AUC was identified where the softmax "probability" of IM was 0.90, with an AUC of .98. In this initial cohort, 1 case was identified as missed IM on a pathologist review (not shown). Recategorization of this case did not materially affect the results.
Full Test Cohort Analysis
To test the behavior of the algorithm and examine its utility in identifying IM in a broad clinical cohort, all cases containing a relevant biopsy protocol in the laboratory management system were identified and final diagnostic text was parsed to separate distinct specimens. Specimens containing keywords (eg, gastric, antrum, antral, stomach) were identified along with their associated specimen diagnosis line. Whole slide images from the initial 2 H&E sections for these specimens were retrieved, where available, and de-identified (4897 total slide images from 2594 blocks; 1935 specimens). Diagnostic text was curated by manual review of every case.
The artificial intelligence (AI) algorithm was applied to the test cohort and specimens were considered "AI positive" if any image from the specimen demonstrated at least 1 patch exceeding the previously fixed 0.90 threshold. Of the 1935 total specimens, 9 (0.5%) demonstrated false-negative results (AI called them negative and pathologist called them positive) and 297 (15.3%) were false positives (AI called them positive and initial clinical pathologist had called them negative).
Specimens with AI discordance with the original pathologist diagnosis were manually examined unblinded to both the original diagnosis and AI diagnosis, including overlays displaying the AI-positive patches as a heatmap (Figure 2, A). Of a total of 513 specimens called positive for IM in the clinical report, the AI algorithm successfully identified areas of IM in 504 (98.2%; Figure 2, B).
Expert review identified areas of obvious AI error, generally false positives. These were commonly in areas not actually containing tissue, which were nevertheless included by the simple Otsu tissue detection method, including areas of bubbles (Supplemental Figure 1, A; see the supplemental digital content containing 2 figures at https://aplm.kglmeridian.com by clicking on the article title in the June 2026 table of contents), squamous debris (Supplemental Figure 1, B), and blood (Supplemental Figure 1, C). Other false-positive detections were present in tissue and included areas of reactive gastropathy and regenerative change (Supplemental Figure 1, D) or parietal cell vacuoles (Supplemental Figure 1, E).
Other common detections in negative specimens included intestinal-type epithelium best explained as "float material" produced as contamination during sectioning (Supplemental Figure 2). In some cases, gastric biopsies included pieces of apparent duodenum, identifiable owing to presence of villi and/or Brunner glands without any clear gastric mucosal background (Figure 3). Together, these truly intestinal-type detections that nonetheless did not represent IM comprised 35 of the 297 apparent false-positive detections (11.7%).
Following manual review, however, some specimens appeared to show true IM that was missed in the initial diagnosis. Therefore all 306 discordant specimens with an admixed 50 concordant positive and 50 concordant negative specimens were reviewed by 3 specialist gastroenterologic pathologists blinded to the initial diagnosis, machine learning result, and ratio of case status types. Specimens were reviewed digitally by using the de-identified slide images in QuPath and scored for presence/absence of IM and on a 4-point score quantitating IM, where present.
Following re-review, 360 of the 406 re-reviewed specimens (88.6%) demonstrated fully concordant results between the initial reading pathologist and the 3 re-reviewers. Fleiss [kappa] score for interobserver concordance was extremely high at 0.824. However, in a small number of cases a separate pathologist unblinded to the AI result disagreed with this concordant diagnosis. For this reason, all cases with any discordance between the initial diagnosis, blinded re-reviewers, and unblinded AI reviewer were discussed in a final consensus conference, where all 4 pathologists were unblinded to the AI findings and the prior pathologist reads. Two cases where consensus could not be achieved were excluded from further analysis.
In all, 30 of the total 1935 specimens (1.6%) were clinically reported as negative but were flagged by AI and ultimately deemed positive by consensus review (Figure 4, A through G). Two cases of the total 1935 (0.1%) were clinically reported positive but deemed negative by the AI algorithm and confirmed negative by consensus review (not shown). The 100 re-reviewed cases without AI disagreement were ultimately declared concordant with the original pathologic diagnosis. The final AUC was again 0.98 after rounding (Figure 5, A).
The number of positive patches identified in clinically missed specimens ranged from 1 to 427 (mean, 20.8; median, 3), and the pathologists' mean semiquantitative scores for IM ranged from 4 (graded as extensive by all re-review readers; Figure 4, A) to 0 (no IM seen by any of the re-review readers; Figure 4, F and G). Only 2 specimens had a mean human quantitation score greater than 2 (focal), and 14 had a mean score greater than 1 (rare goblet cells only). Nineteen specimens with initially missed IM were again missed by at least 1 re-reviewer, and 7 cases were missed by all re-reviewers. Specimens with IM missed by all AI-blinded reviewers most frequently showed only 1 to 3 positive patches by the AI algorithm (6 of 7), though 1 case had 11 positive patches, highlighting the minute nature of these foci.
In re-review of slides and biopsy diagnoses, missed IM was most frequently present in the antrum (16 of 30; 53%) followed by transitional mucosa, defined by histologic co-occurrence of parietal and G cells (7), and oxyntic mucosa (3) or "atrophic mucosa" (4; identified as such in the original specimen diagnosis). Missed IM was overwhelmingly more likely to be complete-type (25 of 30; 83%), with just 3 of the missed IM cases demonstrating incomplete-type. In 2 cases, the type could not be determined with confidence owing to the extremely limited extent of metaplasia.
Missed IM most frequently occurred in a background of gastritis (21 of 30; 70%), with chronic active gastritis present in 9 cases and the remainder showing chronic gastritis without activity. Four cases with active gastritis were positive for Helicobacter pylori. Remaining cases showed either features of reactive gastropathy (5 cases) or were called negative for pathologic changes (4 cases).
To assess the AI's utility for quantitative evaluation, the mean pathologist semiquantitative scores (treated as zero where a pathologist did not identify IM) were compared to the total number of positive patches by the AI. Only cases where both the final consensus was positive for IM and the AI found at least 1 positive patch were considered for this analysis. These demonstrated excellent correlation by Spearman rank correlation coefficient (0.90; P < .001). Positive patches showed an approximately exponentially increasing number with increasing mean score (Figure 5, B).
Llama Interpretation of Diagnostic Text
To evaluate an LLM in the analysis of diagnostic text, an open weights AI model (Llama) was used. (26) Llama showed exemplary concordance with manual review. Overall, 1926 case parts with diagnostic text reviewed (99.3%) showed a conclusive "positive" or "negative" interpretation by the LLM. Remarkably, these showed 100% concordance with manual review. Thirteen cases were flagged by the LLM as equivocal, with 2 reviewed manually as positive and 11 as negative.
Examination of "equivocal" results revealed 2 cases in which there was real ambiguity stated in the diagnosis. One included the phrases "fragments of intestinal type mucosa, which may represent either IM or small bowel mucosa adjacent to the anastomosis." A second was described as "[a] single fragment of intestinal-type mucosa with goblet cells is present. If this originated in the stomach, then this would represent IM. However, the fragment closely resembles duodenum." These had both been labeled as positive in manual review. Four of the other equivocal results included some mention of a comment, but the comment text had not been extracted with the biopsy part text, possibly leading to the equivocal tag. The remaining 7 diagnoses showed no clear reason for the equivocal result.
DISCUSSION
Automated AI review of archival slides enabled identification of IM in 30 of nearly 2000 cases. While this review was performed on de-identified specimens, the demonstrated system could potentially increase sensitivity for IM and enable correction of a missed diagnosis when used as part of a clinical workflow. Importantly, missed IM was found with a variety of background findings, suggesting that general AI-based QA on all nonneoplastic biopsy specimens may be helpful.
While most discordant AI detections represented false positives, 30 of the 297 cases with an AI-positive patch called clinically negative ultimately represented a missed diagnosis (10%). These missed diagnoses represented a spread of both complete- and incomplete-type IM that was variably present in the antral or oxyntic mucosa. While most cases showed some other background pathologic process (commonly gastritis, particularly active gastritis), some cases showed no other pathologic change.
It is uncertain whether the reporting of IM in these specimens would have led to different management in these cases, and in particular increased surveillance. Current American Gastroenterological Association recommendations are conditionally against routine surveillance or repeated short-interval biopsies in response to IM, but recognize that screening may be warranted in patients with incomplete IM, extensive IM (presently defined primarily on the basis of presence of IM in the antrum and corpus rather than extent within a given biopsy), or family history of gastric cancer. (27)
Owing to the fully de-identified nature of the specimens, it is unclear whether IM may have been identified in other specimens in the same case or in prior or subsequent cases. Data regarding follow-up or relevant concurrent clinical diagnoses are likewise not available. However, it is plausible that some cases, particularly those with incomplete IM (3 cases), IM of oxyntic mucosa (5 cases), H pyrlori infection (4 cases), or atrophic changes (4 cases), might have warranted closer endoscopic follow-up had IM been identified, though, again, it is unclear whether IM may have been identified in other biopsy parts of the case or whether biopsies were part of a systematic mapping. Further study is needed to assess the clinical value of using AI to identify IM in a retrospective of prospective manner.
This study additionally provides evidence of using open-source LLMs in the interpretation of pathologist diagnoses. Manual review of surgical pathology reports from thousands of gastric biopsy specimens requires substantial human effort. Simultaneous AI-based screening of both slides and diagnosis for mismatch in an entirely automated fashion would be much less costly in pathologist time. The surprising finding that a "zero-shot" open-source LLM with a relatively simple prompt could handle the vast majority of IM diagnoses with flawless categorization, when conclusive, is an exciting prospect. Importantly, though, this result may not generalize to other diagnoses, nor would it necessarily generalize to other institutions' formatting or phrasing of diagnostic reports. Any extension of this work to another institution, diagnosis, prompt, or underlying LLM should require validation against a large representative sampling of diagnoses with human annotated ground truth, as was used in this work.
In this study, the algorithm was studied in an entirely retrospective context. Use of the algorithm actively in diagnosis as part of a digital sign-out is another possible implementation; however, we did not test that approach in this work. Many implementation paradigms are possible, including display of "heatmap" overlay layers to the signing pathologist, simple flags that the AI has detected some quantity of IM prior to slide review, or preverification checking to ensure that diagnostic text appears concordant with AI interpretation. In this last approach, LLM interpretation of the diagnostic text alongside the image analysis result could ensure that flags are only raised in the event of mismatch.
One advantage of the approach is the low algorithm computational complexity. By operating on the x4 downsampled layer (approximately equivalent resolution to a x10 objective), the algorithm can detect minute foci of IM while saving 16 times the processing power needed for the full-resolution image. Given that many practiced gastrointestinal pathologists operate at x10 (or lower) for most slide reviews, this seems a natural approach.
Despite the abovementioned advantages and significant layers of automation, the described approach suffers from significant false-positive detections (267 of the 297 specimens ultimately reviewed). While total review time was not tracked in this work, it is reasonable to predict that improvements in the algorithm and/or simplifications of the review workflow would be needed for broad adoption. Algorithm improvement could be directed toward initial tissue detection steps to use more sophisticated techniques that exclude areas of blood, bubbles, and stray squamous debris. The neural network algorithm could also be subjected to retraining using the false-positive detections to reduce these false positives. Finally, even without algorithmic improvements, the burden of false-positive review could be mitigated by creating workflows that require only patch-level reviews of the areas of greatest concern found by the algorithm, perhaps followed by consensus review only if a reviewer thought a change in diagnosis might be warranted.
While IM is not currently quantified by pathologists, the algorithm offers an additional feature of potentially giving a quantitative assessment of IM across the biopsy. If normalized by total biopsy area, this might provide a useful measure of the severity of IM within a single biopsy and offer additional useful stratification in addition to anatomic extent. Additional research in cohorts with longitudinal follow-up is ongoing.
While these results demonstrate the potential for using AI as an automated "second reader, " more work is needed to demonstrate extensibility of the findings across institutions. The current algorithm, though trained on data from 2 scanners, is tested only on the output of a single scanner and contains training data from a single institution. Nevertheless, we believe these highlight the potential utility of AI in quality assurance pipelines in the accurate diagnosis of gastrointestinal biopsies.
We would like to acknowledge the participation and resources of the Duke Biorepository and Precision Pathology Center (BRPC), funded through the NCI P30 CA008748. Finally, we also acknowledge the resources of the Duke/Nference collaboration, which created the archival scanned images used in this study.
References
(1.) Sung H, Ferlay J, Siegel RL, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2021 ; 71 (3):209-249. doi:10.3322/caac.21660
(2.) Reddy KM, Chang JI, ShiJM, Wu BU. Risk of gastric cancer among patients with intestinal metaplasia of the stomach in a US integrated health care system. Clin Gas troen terol Hepa tol. 2016; 14(10):1420-1425. doi:10.1016/j .cgh.2016.05.045
(3.) Dixon MF, Genta RM, Yardley JH, Correa P. Classification and grading of gastritis: the updated Sydney System--International Workshop on the Histopathology of Gastritis, Houston 1994. AmJSurg Pathol. 1996; 20(10):1161-1181. doi:10.1097/ 00000478-199610000-00001
(4.) Capelle LG, de Vries AC, Haringsma J, et al. The staging of gastritis with the OLGA system by using intestinal metaplasia as an accurate alternative for atrophic gastritis. Gastrointest Endosc. 2010; 71 (7):1150-1158. doi:10.1016/j.gie.2009.12.029
(5.) Rugge M, Meggio A, Pennelli G, et al. Gastritis staging in clinical practice: the OLGA staging system. Gut. 2007; 56(5):631-636. doi:10.1136/gut.2006.106666
(6.) Lee KJ, Inoue M, Otani T, et al. Gastric cancer screening and subsequent risk of gastric cancer: a large-scale population-based cohort study, with a 13-year follow-up in Japan. Int J Cancer. 2006; 118(9):2315-2321. doi:10.1002/ijc.21664
(7.) Saumoy M, Schneider Y, Shen N, Kahaleh M, Sharaiha RZ, Shah SC. Cost effectiveness of gastric cancer screening according to race and ethnicity. Gastroenterology. 2018; 155(3):648-660. doi:10.1053/j.gastro.2018.05.026
(8.) Morgan DR, Corral JE, Li D, et al. ACG Clinical Guideline: Diagnosis and Management of Gastric Premalignant Conditions. Am J Gastroenterol. 2025; 120(4): 709-737. doi:10.14309/ajg.0000000000003350
(9.) Dinis-Ribeiro M, Libanio D, Uchima H, et al. Management of epithelial precancerous conditions and early neoplasia of the stomach (MAPS III): European Society of Gastrointestinal Endoscopy (ESGE), European Helicobacter and Microbiota Study Group (EHMSG) and European Society of Pathology (ESP) Guideline update 2025. Endoscopy. 2025; 57(5):504-554. doi:10.1055/a-2529-5025
(10.) Song H, Ekheden IG, Zheng Z, Ericsson J, Nyren O, Ye W. Incidence of gastric cancer among patients with gastric precancerous lesions: observational cohort study in a low risk Western population. BMJ. 2015; 351:h3867. doi:10.1136/ bmj.h3867
(11.) Piazuelo MB, Bravo LE, Mera RM, et al. The Colombian Chemoprevention Trial: 20-year follow-up of a cohort of patients with gastric precancerous lesions. Gastroenterology. 2021; 160(4):1106-1117 e3. doi:10.1053/j.gastro.2020.11.017
(12.) Gonzalez CA, Sanz-Anquela JM, Companioni O, et al. Incomplete type of intestinal metaplasia has the highest risk to progress to gastric cancer: results of the Spanish follow-up multicenter study. J Gastroenterol Hepatol. 2016; 31 (5):953-958. doi:10.1111/jgh.13249
(13.) Waddingham W, Nieuwenburg SAV, Carlson S, et al. Recent advances in the detection and management of early gastric cancer and its precursors. Frontline Gastroenterol. 2021 ; 12(4):322-331. doi:10.1136/flgastro-2018-101089
(14.) Coudray N, Ocampo PS, Sakellaropoulos T, et al. Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. NatMed. 2018; 24(10):1559-1567. doi:10.1038/s41591-018-0177-5
(15.) Raciti P, SueJ, Ceballos R, etal. Novel artificial intelligence system increases the detection of prostate cancer in whole slide images of core needle biopsies. Mod Pathol. 2020; 33(10):2058-2066. doi:10.1038/s41379-020-0551-y
(16.) Ehteshami Bejnordi B, Veta M, Johannes van Diest P, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA. 2017; 318(22):2199-2210. doi:10.1001/jama. 2017.14585
(17.) Jiang H, Zhou Y, Lin Y, Chan RCK, Liu J, Chen H. Deep learning for computational cytology: a survey. Med Image Anal. 2023; 84:102691. doi:10.1016/j.media. 2022.102691
(18.) Barmpoutis P, Waddingham W, Yuan J, et al. A digital pathology workflow for the segmentation and classification of gastric glands: study of gastric atrophy and intestinal metaplasia cases. PLoSOne. 2022; 17(12):e0275232. doi:10.1371/ journal.pone.0275232
(19.) Fang S, Liu Z, Qiu Q, et al. Diagnosing and grading gastric atrophy and intestinal metaplasia using semi-supervised deep learning on pathological images: development and validation study. Gastric Cancer. 2024; 27(2):343-354. doi:10.1007/ s10120-023-01451-9
(20.) Epplein M, McCall SJ, Wang F, et al. Prevalence of the cagA virulence factor varies by race among Helicobacter pylori-infected patients undergoing upper endoscopy. Clin Transl Gastroenterol. 2024; 15(6):e1. doi:10.14309/ctg. 0000000000000713
(21.) Bankhead P, Loughrey MB, Fernandez JA, et al. QuPath: open source software for digital pathology image analysis. Sci Rep. 2017; 7(1):16878. doi:10.1038/s41598 017- 17204-5
(22.) Falk T, Mai D, Bensch R, et al. U-Net: deep learning for cell counting, detection, and morphometry. Nat Methods. 2019; 16(1):67-70. doi:10.1038/s41592 018- 0261-2
(23.) He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. Preprint. Posted online December 10, 2015. arXiv:1512.03385. doi:10.48550/arXiv. 1512.03385
(24.) Torchvision maintainers and contributors. TorchVision: PyTorch's Computer Vision library. GitHub. https://github.com/pytorch/vision. Accessed November 21, 2025.
(25.) Otsu N. A threshold selection method from gray-level histograms. Automatica. 1975; 11 (285-296):23-27.
(26.) Grattafiori A, Dubey A, Jauhri A, et al. The Llama 3 herd of models. Preprint. Posted online July 31, 2024. arXiv:2407.21783. doi:10.48550/arXiv.2407.21783
(27.) Gupta S, Li D, El Serag HB, et al. AGA Clinical Practice Guidelines on Management of Gastric Intestinal Metaplasia. Gastroenterology. 2020; 158(3):693-702. doi:10.1053/j.gastro.2019.12.003
John A. Willis, MD; Jacqueline Emerson; Wei Chen, MD; Fengming Chen, MD; Avani A. Pendse, MD, PhD; Katherine S. Garman, MD; William R. Jeck, MD, PhD
Accepted for publication November 21, 2025.
Published online January 22, 2026.
Supplemental digital content is available for this article at https:// aplm.kglmeridian.com by clicking on the article title in the June 2026 table of contents.
From the Department of Pathology, University of North Carolina, Chapel Hill (Willis); Duke University School of Medicine, Durham, North Carolina (Emerson); and the Departments of Pathology (W Chen, F Chen, Pendse, Jeck) and Gastroenterology (Garman), Duke University Hospital, Durham, North Carolina.
This work was funded in part by the Duke University Cancer Disparities P20 from the National Cancer Institute (NCI), P20CA251 657.
The authors have no relevant financial interest in the products or companies described in this article.
Corresponding author: William R. Jeck, MD, PhD, Department of Pathology, Duke University Hospital, Room 327A, Davison Bldg, 40 Duke Medicine Cir, Durham, NC27710 (email: william.jeck@duke.edu).
Caption: Figure 1. Schema of the active learning training process, with initial iterative cooperative automated and manual annotation of intestinal metaplasia and co-training of a UNet segmentation algorithm (A), and subsequent training of a patch-based classifier, using the resulting annotations (B).
Caption: Figure 2. Heatmap-based review of outputs from the patch-based classifier in test-set cases using translucent overlay of the AI output on the whole slide image (A), highlighting positive detection of known intestinal metaplasia at low (B) and high (C) power.
Caption: Figure 3. False-positive detection of intestinal metaplasia via carryover duodenal sampling.
Caption: Figure 4. A through G, Cases of missed intestinal metaplasia with variable extent of missed metaplasia. Short arrows indicate goblet cells in examples (F) and (G).
Caption: Figure 5. A, Overall ROC curve characteristics in test cases following consensus re-review. B, Correlation of mean pathologist IM score in positive cases as compared to number of AI-positive patches. Abbreviations: AI, artificial intelligence; IM, intestinal metaplasia; ROC, receiver operating characteristic.
 Printer friendly
 Cite/link
 Email
 Feedback
| |
| Author: | Willis, John A.; Emerson, Jacqueline; Chen, Wei; Chen, Fengming; Pendse, Avani A.; Garman, Katherine |
|---|---|
| Publication: | Archives of Pathology & Laboratory Medicine |
| Article Type: | Report |
| Geographic Code: | 1USA |
| Date: | Jun 1, 2026 |
| Words: | 5320 |
| Previous Article: | Did I Get It Right? Implementation of Pathology Trainee Report Feedback System Integrated With Digital Pathology. |
| Next Article: | Performance Verification and Establishment of Reference Intervals for Dendritic Cells, [gamma][delta]T Cells, and Follicular Helper T Cells Using... |
| Topics: | |