TBrecon: Towards Integration of DL Image Reconstruction and Post-Processing Tasks
We are currently working on collecting, annotating, and distributing 3D high-resolution knee MRI data, specifically related to knee osteoarthritis. This dataset includes multi-tissue segmentation and anomaly grading. In our project, we aim to leverage the power of multi-task learning to simultaneously address reconstruction and post-processing tasks within an integrated framework, which encompasses segmentation and object detection. To achieve this, we are developing a strategy based on adversarial robust training and confounding factors learning to ensure the reconstruction of small, rare features, even in under-sampled acquisitions.
Related publications
Konovalova N, Tolpadi A, Liu F, Bhattacharjee R, Gassert F, Giesler P, Luitjens J, Majumdar S, Pedoia V. Towards Integrating DL Reconstruction and Diagnosis: Meniscal Anomaly Detection Shows Similar Performance on Reconstructed and Baseline MRI. In Proceedings of the 31st Annual Meeting of ISMRM, Toronto, Ontario, Canada, 2023. 1381.
Project members at UCSF
Natalia Konovalova, Aniket Tolpadi, Peder Larson, Sharmila Majumdar, Valentina Pedoia
Funding number
NIHR01AR078762 (Apr 1, 2021 - Mar 31, 2026)
Principal Investigator
Valentina Pedoia
Reconstruction and Object Detection
Full-Sampled, UNet (L1, SSIM) UNet (L1, K-Space, SSIM) I-Net (L1, K-Space, SSIM)