About The Project

This repository contains the code for the publication titled "A new, machine learning-based approach to metastatic neuroendocrine tumors of unknown origin", accepted by the Journal of Neuroendocrinology.

Installation

1. Clone the repository.

   git clone https://github.com/IPMI-ICNS-UKE/NET_CUP

2. Install the package and necessary dependencies.

   pip install -e .
   pip install -r requirements.txt

3. Download data:
   • Download the required CSV files and feature vector from Zenodo and extract them into the data/ directory.

Additional Steps for Generating New Feature Vectors

To respect data privacy, only feature vectors generated from random patches of whole-slide images (WSIs) are publicly accessible. The original WSIs can be provided upon reasonable request.

1. Download ResNet weights:
   • Download the MTDP ResNet weights from MTDP repository.
   • Download the RetCCL ResNet weights from RetCCL repository under the "Pre-trained models for histopathological image tasks" section.
   • Rename the downloaded files to mtdp.pth and retccl.pth and place them into weights/.
2. Copy WSIs along with their .geojson segmentation files into data/external_dataset/ and data/uke_dataset/.

Usage

Reproducing results

To reproduce the results from the publication, run the Jupyter Notebooks available in src/NET_CUP/experiments. Within each notebook, you can select different pretrained ResNet backbones and classifiers. Configure these settings in the notebook's "Settings" section by changing the classifier variable or setting the feature_type variable to one of the following options as described in the publication:

• FeatureType.IMAGENET
• FeatureType.RETCCL
• FeatureType.MTDP

Using trained models

Different models for binary classification between pancreas and small intestine patches are available in the models/ directory. An ONNX file was created for each combination of pretrained ResNet feature extractor (ImageNet, MTDP, and RetCCL) and classifier (SVC with RBF kernel, SVC with linear kernel, and Logistic Regression). Each ONNX file contains a scikit-learn pipeline that includes a fitted PCA for dimensionality reduction and a trained classifier. To make predictions, load the ONNX file using onnxruntime and run inference with a feature vector:

   import onnxruntime as rt
   sess = rt.InferenceSession(<PATH TO ONNX FILE>, providers=["CPUExecutionProvider"])
   pred = sess.run(None, {"X": <ARRAY WITH FEATURE VECTORS>})[0]

Generating new feature vectors

After acquiring access to the original WSIs and following the Installation steps, you can generate new feature vectors:

python src/NET_CUP/feature_extraction/extract_features.py

License

Distributed under the CC BY-NC-SA 4.0 License. See LICENSE.txt for more information.

Contact

Rüdiger Schmitz - r.schmitz@uke.de. For further details, cf. company and UKE IPMI team pages.

Project Link: https://github.com/IPMI-ICNS-UKE/NET_CUP