This axis deals with the development and validation of methods dedicated for modeling in oncology applications.
The first part of the axis is related to multiparametric predictive models relying on radiomics and other associated data.
Catherine Cheze Le Rest
Reseach activities are carried out related to image analysis and image processing to extract information from multimodal PET/CT and MRI images, also called radiomics. Machine (deep) learning methods are exploited to train and validate models combining radiomics with clinical contextual data and other -omics, to predict various clinical endpoints (diagnosis and virtual biopsies, response to (chemo)radiotherapy, survival, etc.) in a number of cancer types (lung, cervix, esophagus, head and neck, rectum, brain tumors, etc.) and to personalize, optimize and guide treatment strategies.
The developments are carried out by the team (1 post-doc and 4 PhD students) for each of the different "boxes" of the pipeline above:
- Harmonization strategies to enable carrying out robust multicentric studies combining features extracted from images that are produced by different centers.
- Image pre-processing methods such as denoising, interpolation and partial volume effects correction.
- Fully automated detection and segmentation of tumors from PET, CT and MRI.
- Improvement of radiomic features in terms of stability, reproducibility, robustness and discriminative power, as well as development of new original "handcrafted features".
- Using deep neural networks to extract "deep features".
- Comparison of feature selection methods and associated classifiers including deep neural networks, for classification and regression tasks, as well as methods for fusion of several classifiers.
Another part of this axis deals with simulations of tumors at the level of cells populations within the context of radiotherapy and dosimetry.
Contact : Nicolas Boussion