Colorectal cancer is the third most common cancer worldwide, and hereditary polyposis syndromes account for a significant fraction of early-onset cases. Familial adenomatous polyposis, Peutz-Jeghers syndrome, juvenile polyposis syndrome, and ganglioneuroma syndrome each carry distinct genetic signatures and distinct endoscopic appearances, yet distinguishing them from ordinary sporadic polyps during a live colonoscopy remains a challenge even for experienced gastroenterologists. Zahra Ghaffari, Massih Bahar, Mojgan Forootan, Ali Darvishi, and Hamidreza Bolhasani present ERCPMP-Gx, a public dataset designed to give AI systems the multi-modal training signal they need to recognize, characterize, and classify hereditary polyposis at the patient level.

What the Current Landscape Lacks

Public endoscopic datasets for polyp analysis are not hard to find. CVC-ClinicDB offers 612 images from 29 colonoscopy video sequences. HyperKvasir contains over 10,000 labeled GI images. LDPolypVideo-DB provides hundreds of thousands of frames from annotated videos. But all of these are organized around the individual sporadic polyp. They tell you where a polyp is and what it looks like, but they do not connect that polyp to a patient's broader clinical picture. None links the endoscopic phenotype to histopathology and germline findings at the patient level.

This gap matters because hereditary polyposis syndromes are not diagnosed by looking at a single polyp in isolation. A tubular adenoma in a 25-year-old with hundreds of similar polyps across the colon means something entirely different from the same histological finding in a 60-year-old with a solitary lesion. The clinical context, the family history, and the underlying germline mutation all change what the endoscopist should do next. An AI tool trained only on polyp segmentation masks or binary classification labels cannot make these distinctions.

A Three-Layer Annotation Framework

ERCPMP-Gx addresses this by providing a three-layer annotation structure for each record in the dataset. The first layer is standardized endoscopic annotation. Images and video clips are classified using internationally recognized morphological frameworks: the Paris classification for superficial morphology, the Kudo pit pattern classification for crypt architecture, and the JNET (Japan NBI Expert Team) classification for narrow-band imaging assessment of microvascular and surface patterns. These annotations capture what the endoscopist sees during the procedure.

The second layer is histopathology. Each record is linked to a representative pathological diagnosis confirmed by board-certified gastrointestinal pathologists. Diagnoses include tubular adenoma, villous adenoma, tubulovillous adenoma, hyperplastic polyps, serrated lesions, inflammatory polyps, and adenocarcinoma, along with dysplasia grade and differentiation grade. This provides the ground truth for what the polyp actually is at the tissue level.

The third layer, and the one that distinguishes ERCPMP-Gx from every existing public dataset, is the genomic component. Where available, each record is linked to clinically reported germline findings. For familial adenomatous polyposis, this means APC gene mutations. For Peutz-Jeghers syndrome, STK11/LKB1 mutations. For juvenile polyposis syndrome, BMPR1A or SMAD4 mutations. For ganglioneuroma syndrome, RET proto-oncogene mutations. This three-way linkage between endoscopic appearance, histopathology, and genotype at the patient level is what makes the dataset genuinely AI-ready for precision medicine applications.

Imaging Under the Olympus EVIS X1

All procedures in the dataset were performed using the Olympus EVIS X1 system, which is the most recent endoscopy platform from Olympus. The dataset includes images captured under four distinct imaging modes. White-light endoscopy provides standard mucosal visualization. Narrow-band imaging uses 415nm and 540nm wavelength filters to enhance the contrast of mucosal surface patterns and submucosal vascular architecture, which is critical for assessing pit patterns and microvascular changes associated with dysplasia. Magnifying NBI combines optical zoom with narrow-band filtering to enable detailed assessment of individual pit structures and capillary networks. Near-focus mode brings the working distance down to 2-6mm, providing close-up observation of mucosal surface detail that is essential for differentiating hamartomatous from adenomatous tissue.

The result is 160 high-resolution RGB colonoscopy images with accompanying video clips. The dataset is released at the patient level, with each record containing the image or video, its endoscopic annotation, its histopathological diagnosis, and its germline findings where available. The dataset is publicly accessible on Mendeley Data under a CC BY-NC 4.0 license, with progressive releases planned as additional video content becomes available.

Why 80/20 Is the Right Split

Approximately 80% of the cases in ERCPMP-Gx represent clinically and/or genetically confirmed hereditary polyposis syndromes. The remaining 20% comprise non-hereditary polyps and polyp-mimicking lesions with overlapping morphological features. This is not a convenience sample. The non-hereditary cases are included deliberately to support differential classification.

During a live colonoscopy, the endoscopist does not know in advance whether a given polyp is hereditary or sporadic. The visual appearance of a juvenile polyp in Peutz-Jeghers syndrome can overlap with a hyperplastic polyp in a patient with no family history. An AI system that has only seen hereditary polyposis cases will over-diagnose, while one that has never seen them will miss them entirely. The 20% non-hereditary fraction provides the hard negative examples that force the model to learn the actual distinguishing features rather than defaulting to a majority class prediction.

The Clinical Stakes

Hereditary polyposis syndromes are not just a curiosity of medical genetics. They are precancerous conditions with real consequences for patients and their families. Familial adenomatous polyposis, left untreated, progresses to colorectal cancer with near certainty. Peutz-Jeghers syndrome carries elevated risk of multiple gastrointestinal and extra-gastrointestinal cancers. Juvenile polyposis syndrome increases colorectal and gastric cancer risk. Ganglioneuroma syndrome is associated with medullary thyroid carcinoma through MEN2B.

Early identification of these syndromes changes clinical management fundamentally. A patient diagnosed with FAP in their twenties will undergo prophylactic colectomy and enter a lifelong surveillance program. Their first-degree relatives will be offered genetic testing. The clinical trajectory for the entire family shifts. But these diagnoses depend on the endoscopist recognizing that the pattern of polyps they are seeing is not ordinary. When the endoscopist misses the pattern, or when the procedure is performed by a less experienced clinician, the diagnosis is delayed.

An AI system trained on ERCPMP-Gx could serve as a real-time decision support tool, flagging cases where the endoscopic pattern is consistent with a hereditary polyposis syndrome and recommending genetic testing. This is particularly valuable in settings where access to experienced gastroenterologists with expertise in hereditary cancer syndromes is limited.

How ERCPMP-Gx Compares

The dataset occupies a unique position in the landscape of public colonoscopy datasets. CVC-ClinicDB and ETIS-Larib focus on polyp detection and segmentation without histopathological detail. HyperKvasir and LDPolypVideo-DB provide large volumes of images but are organized around general GI endoscopy rather than hereditary conditions. PICCOLO offers bounding boxes and labels for polyp detection but does not integrate germline data.

ERCPMP-Gx is smaller in raw image count than some of these datasets, but it is substantially richer in annotation depth. The combination of standardized endoscopic classification, confirmed histopathology, and germline findings creates a training signal that no existing dataset provides. For researchers building AI tools that go beyond polyp detection toward clinical decision support for hereditary cancer syndromes, this is the first public resource that supports the full diagnostic pipeline from endoscopic appearance to genetic characterization.

Limitations and Next Steps

The dataset's 160 images represent a focused release from what the authors describe as a broader ERCPMP project. The relatively small size is a constraint for training large deep learning models, though the richness of the per-image annotation partially compensates. Video content is being released progressively in subsequent versions, which will enable temporal analysis of polyp appearance across a procedure.

The dataset draws from two Iranian medical centers (Shahid Beheshti University of Medical Sciences in Tehran and Shiraz University of Medical Sciences), which means the patient population and the spectrum of hereditary syndromes may not be representative of all clinical settings. The germline findings are described as "where available," indicating that not every record has complete genomic data, which introduces some incompleteness in the genotype layer.

The CC BY-NC 4.0 license restricts commercial use, which may limit adoption in industry settings but is appropriate for a dataset containing sensitive clinical information. The DataBioX website (databiox.com) is maintained as the central point for updates and additional resources.

What This Means for AI in Gastroenterology

ERCPMP-Gx is a dataset paper, not a methods paper. It does not propose a new architecture or report state-of-the-art benchmarks. Its contribution is providing the data that enables other researchers to build and evaluate AI systems for a clinical problem that existing datasets cannot address. For the computer vision and medical AI communities, the value lies in the annotation structure: a patient-level framework that connects what the endoscope sees, what the pathologist confirms, and what the geneticist finds.

This kind of multi-modal, patient-level annotation is what separates a useful clinical AI tool from a research curiosity. Polyp detection is a solved problem in the sense that multiple systems achieve high sensitivity and specificity for finding polyps in colonoscopy video. The next frontier is characterization: telling the endoscopist not just that a polyp exists, but what it means for this patient, in this family, given this genetic context. ERCPMP-Gx provides the training data to make that possible for hereditary polyposis syndromes for the first time in a public resource.

Read the paper on arXiv