Project realised during a 6 months internship at IUCT Oncopole, France.
This provides some deep Learning tools for automatic segmentation of medical images. The approach implemented for this project is to process the entire medical acquisition at the same time. One of the major challenges when processing this kind of data using deep learning algorithms is the memory usage, as depending on the modality and the study, an imaging serie can contains several hundreds or thousands of images.
The model used during this project is a custom U-Net [1], adapted to handle 3D medical images, reduce overfitting and limit the RAM comsumption. The model V-Net [2] is also implemented but couldn't be used, as it is more elaborate and requires more RAM, which wasn't possible.
[1] O. Ronneberger, P. Fischer, and T. Brox, ‘U-Net: Convolutional Networks for Biomedical Image Segmentation’, arXiv:1505.04597 [cs], May 2015.
[2] F. Milletari, N. Navab, and S.-A. Ahmadi, ‘V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation’, arXiv:1606.04797 [cs], Jun. 2016.
-
Tissue segmentation of CT scan in 5 classes:
Background / Fat / Soft tissues / Lungs / Bones.results: no signs of overfitting, visually correct, median DSC > 0.9
-
Physiologic segmentation of IRM in 9 classes:
N/A / Spleen / Liver / 6 lymphatic nodes.results: signs of overfitting, visually correct on Liver and Spleen, wrong for lymphatic nodes, median DSC < 0.4
-
Tumour segmentation of PET scan.
results: signs of slight overfitting, visually acceptable, median DSC > 0.65
-
Tumour segmentation of PET/CT scan.
results: no signs of overfitting, visually correct, median DSC > 0.74
Trained model available: /master/deeplearning_models/trained_model_09241142.h5
- modality PET-CT available
/master/class_modalities/modality_PETCT.py
- fix image shifting between PET and CT
- add modality CT
- add modality IRM
- add modality PET