Loading...
Loading...
Contributors:
Dataset URI:
10.5281/zenodo.23175774Model weights of the histology classifier used in ColoRisk, a research prototype for the post-procedural review of recorded colonoscopies developed by Team NU-Kazakhstan (Nazarbayev University) for iGEM 2026. The source code is in the linked GitLab repository, and the team wiki (https://2026.igem.wiki/nu-kazakhstan/) describes the project in detail. Authorship and AI assistance. The model was trained by Aruzhan Sabrgaliyeva. An AI assistant (Claude, by Anthropic) helped with code, debugging and documentation. The author is responsible for the content of this record. Not a diagnostic device. The model is intended for research and education. It has not been clinically validated, and any finding must be verified by a physician. What the file is. A PyTorch state_dict of the timm model efficientnet_b3 with two output classes, in this order: adenoma, hyperplastic. Input: an RGB crop of a polyp bounding box, resized to 224 x 224 pixels and normalised with the ImageNet mean and standard deviation. Training. Starting point: ImageNet-pretrained timm weights. The last blocks (blocks.5, blocks.6), the head convolution, the final batch normalisation and the classifier were fine-tuned, and the rest of the network was frozen. 40 epochs, batch size 32, AdamW (learning rate 5e-5, weight decay 1e-4), cosine annealing, class-weighted cross-entropy, random flips, rotation up to 15 degrees, colour jitter and small translations. The checkpoint with the best validation macro-F1 was kept. Data. KUMC PolypsSet (Li et al., PLoS ONE 2021, doi:10.1371/journal.pone.0255809; data on Harvard Dataverse, doi:10.7910/DVN/FCBUOR). According to the dataset paper, it combines four sources (MICCAI 2017, CVC-ColonDB, GLRC and KUMC) into 155 video sequences with 37,899 frames, split at sequence level into 116 training, 17 validation and 22 test sequences. Patches were cut from the annotated bounding boxes and labelled adenomatous or hyperplastic from the XML annotations (mapped to adenoma and hyperplastic). Four cases were excluded from training and validation and used only for the final evaluation. The remaining 34,564 patches were split at random 85:15 into 29,380 training and 5,184 validation patches. The data itself is not included in this record. Evaluation. On the four held-out cases (1,417 patches: 703 adenoma and 714 hyperplastic): accuracy 0.90. Adenoma: precision 0.88, recall 0.93, F1 0.91. Hyperplastic: precision 0.93, recall 0.88, F1 0.90. Confusion matrix (rows: true, columns: predicted; adenoma first): [[652, 51], [85, 629]]. Other ColoRisk models. The weights of the ColoRisk polyp detector are not part of this record. The repository README explains why. Known limitations. Only two classes. The dataset has no serrated lesions, so the model cannot recognise them and always returns "adenoma" or "hyperplastic" for any input. The train and validation split is at patch level, so patches of the same patient can appear in both. The checkpoint was selected on this validation split, and its scores (macro-F1 of about 0.99) are optimistic. Use the held-out result above. Evaluated on one dataset only (KUMC PolypsSet). Performance on other endoscopes, hospitals and populations was not tested. Patient-level separation cannot be verified for the part of the training data that comes from the train2019 folder of the dataset, which has no case structure (the dataset authors shuffled its frames into one folder). How to load. import timm, torch model = timm.create_model("efficientnet_b3", pretrained=False, num_classes=2) model.load_state_dict(torch.load("histology_classifier.pt", map_location="cpu")) model.eval() # class order: ["adenoma", "hyperplastic"] Tested with timm 1.0.30 and PyTorch 2.14.0 (the versions used for the ColoRisk web application). Terms of use. The software in the linked repository is licensed under Apache 2.0. These weights were trained on the dataset described above, which its authors released under CC0 1.0 (Harvard Dataverse, doi:10.7910/DVN/FCBUOR) and ask users to cite. The weights are shared under CC BY 4.0, with attribution to this record and to the dataset authors. The dataset combines four earlier public datasets (MICCAI 2017, CVC-ColonDB, GLRC and KUMC), so check the terms of those sources before any commercial use. The starting weights come from timm (ImageNet-pretrained efficientnet_b3), and their terms are on the timm model card. This model is not a diagnostic device.
Keyword(s):
2026; adenoma; colonoscopy; deep learning; EfficientNet.
MeSH Term(s):
Grants and Funding:
No Grants or Funding Data Available
Associated Publication(s):
No Associated Publication Data Available
Supplementary Data:
No Supplementary Data Available
Related Dataset(s):
No Related Dataset Data Available
The National Library of Medicine (NLM) provides information about biomedical datasets. Inclusion on an NLM Web page does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: NLM Web Policies (nih.gov)