Benoit Freiche

Postdoctoral fellowship on AI for medical imaging. Research on Ultrasound image generation. Design of new methods using Stable Diffusion and ControlNet for the generation of Ultrasound image of breast cancer, thyroid cancer and heart disease. Design of a method for ultrasound image segmentation in low-data regime using generative AI (text-to-image models). Implementation with Python (Pytorch, versioning with GitHub), publication of the results, and presentation at an international conference in February 2025 (SPIE medical imaging). Work on quantitative ultrasound (Physic-based optimization in Matlab, prediction of the diffuser density in ultrasound images). Study for Siemens. Ultrasound and RF data collection at the hospital (Universitary center of McGill) for the study of breast cancer-related lymphedema. Supervision of three students on various deep and machine learning topics. Consistent review of deep and machine learning litterature, presentations to the team. Popularisation of science at various Concordia university events.

PhD (machine learning/computer vision/signal processing) defended on the 7th of December 2024: Hierarchical integration of cardiac imaging data with machine learning. Development and implementation of two machine learning methods for the hierarchical integration of multi-modal imaging data in the context of acute myocardial infarction. The first method is based on manifold learning and was presented in two international conferences (STACOM 2021 and MAI 2021). The second method is based on Gaussian Processes Latent Variable Models and was published in IEEE Transactions on Medical Imaging (TMI), a top international peer-reviewed journal (Code available at https://github.com/creatis-myriad/HGP_LVM). Main topics: unsupervised learning, feature engineering, cardiac medical imaging (MRI), hierarchical data integration, multimodal learning, data vizualisation.

Deep Reinforcement Learning internship for the early prediction of bacteria resistance, on time series of images. Design of a solution for the scaling from 2D to 3D of a network, through preprocessing of image features and desitgn of a new network architecture using depthwise convolutions. Implementation of various deep learning architectures, Inception blocks, depthwise convolutions, variational auto-encoders (VAEs). Versioning with Git, distributed computing on a cluster of GPUs. Presentation and demonstration of VAEs at team meetings.

R package implementation with Rcpp, linked to a publication in extreme statistics domain.