Description

A pre-trained model for interactive segmentation of organs/lesions in 2D medical images/slices.

Model Overview

This semi-automatic segmentation method is trained using the framework introduced in [1] that leverages a Unet-like fully convolutional neural network (FCNN). The algorithm (i) responds to a simple form of interaction by the user, more specifically mouse clicks (foreground/background); (ii) is able to deliver accurate results with the smallest amount of interaction (1 click), but allows for refinement of the results up to arbitrary precision by moving the click position or adding additional interactions. In other words, as the number of interactions grows, performance grows accordingly; (iii) generalizes and performs well on previously unseen structures.

Data

The model is trained using both Synapse Multi-atlas labeling dataset (DOI: 10.7303/syn3193805) and Medical Segmentation Decathlon (MSD, http://medicaldecathlon.com/) dataset as the training data. Multiple organs/lesions are included, i.e., adrenal glands, aorta, esophagus, gall bladder, kidneys, liver, pancreas, splenic/portal veins, spleen, stomach, and vena cava from Synapse, and lung tumor, liver and tumor, spleen, pancreas and tumor, colon cancer, heart, prostate and tumor from MSD.

Training configuration

Currently, the training functionality of this model is not available on the Clara Train SDK.

Input and output formats

Input: 1 channel MR/CT image/slice + foreground and background clicks Output: 2 channels: Label 1: target; Label 0: everything else

Scores

This model achieve the following Dice score on the validation data (our own split from the training dataset of MSD):

1. CT Spleen: 0.909 (1-click) 0.966 (5-click)
2. CT Liver and lesion: 0.785 (1-click) 0.942 (5-click)
3. CT Pancreas and lesion: 0.667 (1-click) 0.893 (5-click)
4. CT Colon lesion: 0.659　(1-click) 0.9 (5-click)
5. CT Lung lesion:　0.772 (1-click) 0.911 (5-click)
6. MR Heart: 0.781 (1-click) 0.942 (5-click)
7. MR prostate and lesion: 0.736 (1-click) 0.873 (5-click)

Availability

This model is designed to be used in AIAA. You can download the model from NGC registry as described in Getting Started Guide

Disclaimer

This is an example, not to be used for diagnostic purposes

License

End User License Agreement is included with the product. Licenses are also available along with the model application zip file. By pulling and using the Clara Train SDK container and downloading models, you accept the terms and conditions of these licenses.

References

[1] Sakinis, Tomas, et al. "Interactive segmentation of medical images through fully convolutional neural networks." arXiv preprint arXiv:1903.08205 (2019).