FAR-POLYP-SEG: A Prospective Single-Center Colonoscopy Dataset for Colorectal Polyp Segmentation with Patient-Level Metadata and Baseline Cross-Dataset Evaluation

Reliable computer-aided colonoscopy depends on well-organized imaging data, transparent benchmarking, and validation of artificial intelligence (AI) tools on independent data. Many public colonoscopy datasets are limited in size, omit normal-mucosa frames, or lack linked patient-level clinical context, which constrains their use as imaging informatics benchmarks. We present FAR-POLYP-SEG, a prospective single-center colonoscopy dataset for colorectal polyp segmentation, developed as an imaging informatics resource for the validation of segmentation algorithms. The dataset was acquired at Farhikhtegan Hospital, Islamic Azad University, Tehran, Iran, between February and December 2025 during routine colonoscopy, without modification of the standard diagnostic workflow. It contains 8181 frames from 455 patients: 432 polyp-positive frames with expert pixel-level segmentation masks and 7749 normal-mucosa frames. Patient-level clinical and procedural metadata (age, sex, colonoscopy indication, Boston Bowel Preparation Scale (BBPS) score, and procedure duration) are linked to every case. Six segmentation architectures (UNet, UNet++, UNet (MiT-B0), nnU-Net 2D, PraNet, and YOLOv11m-seg) were trained and evaluated under one standardized protocol using patient-grouped five-fold cross-validation, so that no patient contributed frames to both the training and the test set of any fold. PraNet reached the highest internal segmentation performance (Dice 0.755), while nnU-Net reached the highest internal IoU (0.665) and pixel accuracy. A complementary evaluation on the dataset's 7749 normal-mucosa frames shows that the models most sensitive to real polyps are not the most specific: gated false-positive rates on normal mucosa ranged from 24.9% (YOLOv11m-seg) to 59.5% (nnU-Net), an almost complete inversion of the internal Dice ranking. All models were then evaluated on the public Kvasir-SEG dataset as an unseen external test set; Dice ranged from 0.756 to 0.829 across models, and the relative ordering was largely preserved, although nnU-Net's internal advantage on accuracy and IoU did not carry over to external Dice. The dataset, structured metadata, and evaluation code are publicly released to support reproducible imaging informatics research.