Castelnau-le-Lez, Occitanie, France
Data Scientist with expertise in Computer Vision and Deep Learning. I have developed innovative solutions in Artificial Intelligence for medical imaging, such as training segmentation algorithms on rare medical images, optimizing deep learning algorithms for faster CPU inference, or designing a new neural network architecture for deformable registration of 3D volumes with competitive inference timing on GPU. I am looking for new opportunities to apply artificial intelligence to medical imaging.
Non-rigid registration of 3D medical images: - Selected two deep learning models from the literature. - Optimized the unsupervised registration training with a learnable loss function. - Evaluated the trained deep learning models against the state-of-the-art (SOTA) non-rigid registration method. - Designed a non-rigid registration neural network (SiReDiReg) with competitive registration performances and 10 sec inference timing on GPU compared to 1 min 20 sec for the SOTA method. - Mentoring an intern on the sternum segmentation project. Liver ablation segmentation integration: - Collaborated with a software engineer to deliver the ablation segmentation algorithm with the onnx-runtime as a Windows executable for integration into the Epione software. - Lead the statistical analysis of the ablation segmentation Dice score (79%) validation on Epione's interventional cases, which includes new artefacts and treatments not present in the training dataset.
Prométhée project: Liver and tumor segmentation (10 months). - Delivered a liver segmentation algorithm with 96% Dice score on the LiTS benchmark (7th) and 51 sec mean execution time on CPU per contrast-enhanced CT volume. - Delivered a Dockerized algorithm and a Slicer3D demonstrator for stakeholders. - Mentoring an intern on tumor segmentation. Tumor and liver ablation segmentation project (4 months). - Trained two segmentation models on 95 tumors and 31 ablation cases. - Improved the segmentation Dice score with data augmentation by 2% and 18%, respectively. - Mentoring an intern on data augmentation. Improving liver ablation segmentation project (6 months). - Selected an optimized neural network architecture for CPU execution (RITnet) - Annotated 146 liver ablation cases. - Improved training with synthetic data, which is responsible for 8% of the final Dice score. - Delivered a segmentation algorithm with a mean execution time of 550 ms per ablation and 81% mean Dice score, beating SOTA by 3%. - Mentoring an intern on neural network architecture selection. - Wrote a patent on training a segmentation algorithm on synthetic data from radiologists' manual segmentations. Improving liver tumor segmentation project (6 months). - Participated in the tumor annotation campaign of an additional 1092 tumors. - Improved training with synthetic data, amounting to 3M€ in data budget savings. - Improved training regularisation with semi-supervised training on unannotated data. - Delivered a tumor segmentation algorithm with a 75% Dice score on metasases in the liver and a 73% Dice score on liver primary tumors, beating SOTA by 2% and 9%, respectively.
Projects and technical stack : -- Artificial Intelligence Insights into Regulations, Canadian School of Public Service -- [scikit-learn, Neo4J, Linkurious] Insights into Canadian regulations with graph analytics and natural language processing (NLP). -- Testing & AI -- Industrialization and tests for AI projects. -- Sales predictions -- [Python, Predictive Objects (TellMePlus)] Comparative analysis of the current forecast model to machine learning models from Predictive Objects (TellMePlus) for a retail store. We achieved an impressive 40% decrease in prediction error! -- Predictive maintenance -- [nfif, Kafka, Spark streaming, Scala] Real-time predictions of HDD breakdown from a cloud provider (Backblaze). -- Predictive support -- [Python, Scikit-learn, Django, Heroku] Support ticket classification by machine learning and NLP in three languages. -- Fraud detection -- [Python, Scikit-learn, Neo4j, Linkurious] Fraud detection framework by machine learning with a visualization interface allowing inspectors' annotations to further decrease detection errors. -- Gas price predictions -- [Spark notebook, Scala] Spark demo for parallel computing of gas prices in the country's stations from open-data (https://www.prix-carburants.gouv.fr/). -- GraphGist challenge -- [Neo4j, Cypher] Build Star Wars models with your Lego collection (Holidays GraphGist award from Neo4j) -- CSE NORD, Société Générale -- [SAS, MVS, MicroStrategy, Teradata, DataStage]
I am conducting high-throughput sequencing data analyses and manage several bioinformatics projects in genome regulation, both as a leader or in collaboration.