About me

Automated assessment of interstitial lung disease (ILD) in chest CT imaging @ Brainomix (Oxford, 2022–current)

I developed and deployed an explainable respiratory phase classifier for chest CT using shape features derived from deep learning-based trachea and lung segmentations to support the reliability of in-house ILD biomarkers. I also conducted survival analysis using random survival forests and Cox proportional hazards models to assess the predictive power of these biomarkers. I automated this analysis and optimized memory and speed to improve ML testing workflows in CI/CD. Finally, I worked on denoising as a preprocessing step (e.g., via guided filtering) and level-set-based postprocessing to enhance airway and lung segmentation.

Deep generative modeling and NLP for finance @ Sony Computer Science Laboratories (Tokyo, 2022)

I researched NLP algorithms, including approaches involving transformer-based language models (BERT and MPNet) and LSTMs with GloVe embeddings, for problems such as fake-news and greenwashing detection, supporting Nomura Asset Management in assessing corporate sustainability. I also developed a method inspired by the optical-flow algorithm to estimate time-dependent loadings in security factor models and trained a variational autoencoder to learn and visualize low-dimensional fund-month exposure embeddings, supporting investment analysis at GPIF, one of the world's largest pension funds.

• Project Webpage

Cine-MR image forecasting with RNNs trained online and transformers for latency compensation in radiotherapy @ The University of Tokyo (2017–2021, independent work until 2026)

I designed an algorithm to forecast frames in chest and liver MR image sequences for safe MR-guided radiotherapy, combining deformable image registration, PCA-based motion modeling, and temporal dynamics prediction with RNNs trained online and transformer encoders. The proposed approach is modular, interpretable, and data-efficient. It can help overcome motion irregularity and distributional shift between the training and test datasets for population models.

• YouTube (English sub, 450k views)
• ArXiv
• Code

Marker-position forecasting using RNNs trained online and transformers for safer radiotherapy @ The University of Tokyo (2020–2021, independent work until 2026)

I derived implementations of online training algorithms for RNNs (UORO, SnAp-1, and DNI) with closed-form simplifications, improved accuracy, and reduced asymptotic time/memory complexity. This study is the first to assess their ability to forecast the 3D positions of external markers on the chest and abdomen for latency compensation and robotic control in radiotherapy. Experimentally, modern online-learning algorithms enabled RNNs to adapt fairly well to irregular breathing patterns and achieve competitive accuracy with less data than prior deep-learning approaches, while keeping processing time relatively low.

UORO article:

• Publication
• ArXiv
• Code
• Blog article

Follow-up article (covering also SnAp-1 and DNI):

• Publication
• ArXiv
• Code

Simultaneous segmentation and recognition of characters @ Idemia (Île-de-France, 2016)

I prototyped an OCR pipeline in C for degraded text images, consisting of three successive steps: over-segmentation, character recognition, and shortest-path graph decoding using Viterbi's algorithm to select the most likely string (see the figure above). It achieved 13.4% higher accuracy than Tesseract 3.04, an open-source OCR engine. Over-segmentation (i.e., candidate cut generation) relied on Dijkstra's algorithm applied to a graph representation of the input image, where each pixel was assigned to a node and edge weights were based on pixel intensities and heuristics; I optimized asymptotic time complexity using a binary-search-style strategy. For the second step in the pipeline, I implemented from scratch a character-level neural-network classifier, trained it on ~20k character images, and improved accuracy by ~5% by incorporating a class-conditional Bayesian aspect-ratio prior.

Graph-structured data classification for computer-aided diagnosis in neurology @ Fresnel Institute (Marseille, 2016)

I implemented a binary classifier for graph-structured data, combining generalized likelihood ratio testing and Fourier graph transforms, based on prior work from Hu et al. I applied this method to Alzheimer's disease detection, modeling brain regions in PET images as nodes in a graph with edge weights defined by Gaussian RBF kernels over regional imaging features, and obtained a leave-one-out test F1 score of 0.85 on a dataset of 142 brain scans. Last, I explored optimization of region-independent weights corresponding to each image property to maximize classification accuracy.

• Blog article
• Code

Participation in the French robotics cup (2015)

I built and integrated two autonomous robots that execute actions on a table to score as many points as possible during 90-second matches, as part of a team of 6 students. Specifically, I programmed Arduino microcontrollers in C/C++, focusing on motion control (via a proportional controller), obstacle avoidance, and action sequencing. I also took part in project planning, communications, fundraising/partnerships, and mentoring junior students.

Mechanical modeling of a periodic bundle of fibers @ IEMN institute/LIMMS (Hauts-de-France, 2015)

I modeled a periodic inhomogeneous fiber-bundle structure using coupled ordinary differential equations (ODEs) to study effective stiffness under different conditions. This supports our understanding of DNA degradation mechanisms due to irradiation, as part of the SMMIL-E project, which promotes the development and use of BioMEMS against cancer. I estimated numerically the effective Young's modulus for chains of n repeated unit cells with a transfer-matrix formulation using MATLAB, and analyzed sensitivity to structural parameters. I also derived closed-form expressions for the effective modulus in simplified configurations, and characterized asymptotic regimes (weak/strong coupling and large-n behavior).

• Report (in French)
• Code (with English Readme)

Fluid flow modeling for the development of a non-contact borehole vibration sensor @ Schlumberger (Kanagawa, 2014)

I developed a fluid-flow model (based on Navier-Stokes equations) to assess the feasibility of a non-contact borehole vibration sensor for geological formation analysis, requiring less time and energy to operate than conventional contact-based devices. I considered different geometry configurations and fluid types (e.g., non-Newtonian Bingham fluids) and estimated key physical parameters numerically with MATLAB. I derived and interpreted analytical pressure and velocity expressions as functions of wellbore vibration parameters using partial differential equations (including diffusion-type equations), Fourier analysis, and asymptotic expansions, to further validate sensor feasibility and back up design recommendations.