Conference
PACER: Curriculum Self-Training with Spatial Uncertainty for Pseudo-Label Selection in Few-Shot Anatomical Landmark Detection
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)
Loading...
Loading...
Developing accurate, efficient, and fair AI systems for medical imaging applications.
Studying uncertainty estimation, robustness, and interpretability of medical AI, including 3D diffusion models for counterfactual brain-image generation and sparse autoencoders to explain how counterfactuals are produced.
Accurate identification of anatomical keypoints in diagnostic images (MRI, X-ray), crucial for surgical planning, clinical support, and advanced image processing. Automation via deep learning enables fast, reliable analysis.
Annotated medical data are scarce because of privacy, expert-labeling needs, and time/cost constraints. Using transfer learning, self-supervised learning, and generative models, the goal is to reach state-of-the-art results in low-data regimes.
Investigating disparities in landmark detection across demographic attributes and using diffusion models to mitigate hidden dataset biases — towards fairer, more reliable AI systems whose decisions are not driven by spurious correlations.
Academic partnerships and projects
Télécom Paris, Institut Polytechnique de Paris — under the supervision of Prof. Pietro Gori
Research visit under Prof. Pietro Gori, in close collaboration with his PhD student Akshita Kumar, building on recent work developed within the team.
Technologies
12 peer-reviewed papers
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)
Medical Image Understanding and Analysis (MIUA), Dublin
Third Prize, Best Poster
BMC Bioinformatics, 27, Article 141
International Conference on Image Analysis and Processing (ICIAP), LNCS vol. 16168, pp. 127–139
Veterinary Sciences, 12(5), 404
Smart Tools and Applications in Graphics (STAG)
IEEE International Symposium on Biomedical Imaging (ISBI), pp. 1–5
Awards, oral presentations and conferences
Paper selected among the top 2% of 4,600 submissions and presented to about 1,000 attendees.
Awarded at MIUA 2026 for “Six Landmarks, One Score: Automated Feline Vertebral Heart Scale from Thoracic Radiographs”.
Oral and poster presentation of CDPM-Align, selected from 4,600 submissions (top 2%) and presented to about 1,000 attendees.
Poster on automated feline vertebral heart scale from thoracic radiographs, awarded Third Prize for Best Poster Presentation.
International Computer Vision Summer School: lectures by researchers from Stanford, MIT, Oxford, and Meta Reality Labs on 3D vision, foundation models, generative AI, and spatial intelligence.
Presented diffusion-based self-supervised pre-training for few-shot anatomical landmark detection in X-ray images.
Supported logistics and coordination for over 6,700 participants and 2,300 paper presentations.
Presented transfer learning for landmark detection in small, domain-specific datasets, showing that ImageNet pre-training can be sufficient for high accuracy.
Large-scale AI training, multi-GPU optimization, and scalability challenges in deep learning.
Peer review activities