HoliBrain

ANR-23-CE45-0020-01 (2024-2028)

Abstract

When a complex system is studied, it is usually broken down into smaller and simpler subsystems in order to facilitate its analysis. Such a reductionist paradigm is powerful; making the analysis easier, but ignores the interactions and relationships that exist between the parts and scales of the whole system. Faced with the complexity of brain anatomy, human experts (e.g., neuroanatomists) follow this paradigm when analyzing the cerebral organization using 3D Magnetic Resonance Imaging (MRI). Therefore, human experts specialize in specific brain structures (i.e., task) at a given anatomical level (i.e., scale). This specialization is accentuated at the substructure level – the analysis of which is now possible with high- resolution multimodal MRI – because more information needs to be processed simultaneously. This reductionist strategy results in numerous task-specific image processing methods ignoring task relationships and scale interdependencies. Consequently, these tools produce partial brain representations, preventing global analysis of the brain. HoliBrain introduces a shift towards a novel holistic paradigm able to process high-resolution multimodal 3D MRI in order to generate whole brain analysis a consistent manner. Previously beyond the capabilities of both computational tools and human experts, the recent advances of deep learning open up new opportunities to tackle this great challenge. To this end, HoliBrain proposes to develop new image processing methods based on a large number of networks sharing information between tasks and scales. The main idea is to divide this intractable problem for a single network into multiple simpler tasks processed by a large ensemble of collaborating networks. HoliBrain will produce novel knowledge on cerebral organization and new biomarkers for early diagnostic of neurological. 

Consortium

• • Pierrick Coupé (PI, CNRS, LaBRI UMR 5800 / Bordeaux, France): Medical Imaging

  • Michael Clément (ENSEIRB, LaBRI UMR 5800 / Bordeaux, France): Deep Learning

• • Boris Mansencal (ENSEIRB, LaBRI UMR 5800 / Bordeaux, France): Computer science

  • Floreal Morandat (ENSEIRB, LaBRI UMR 5800 / Bordeaux, France): Cloud-based platform

  • Rémi Giraud (ENSEIRB, IMS UMR 5218 / Bordeaux, France): Image Processing

  • Baudouin de Senneville (CNRS, IMB UMR 5251 / Bordeaux, France): Mathematics

  • Nicolas Papadakis (CNRS, IMB UMR 5251 / Bordeaux, France): Mathematics

  • Thomas Tourdias (INSERM, Neurocentre Magendie U 1215 / Bordeaux, France): Neuroanatomy

  • Aurélie Ruet  INSERM, Neurocentre Magendie U 1215 / Bordeaux, France): Neurology

  • Bruno Brochet  (INSERM, Neurocentre Magendie U 1215 / Bordeaux, France): Neurology

  • Fanny Munsch (INSERM, Neurocentre Magendie U 1215 / Bordeaux, France): Medical IMaging

  • José Manjon (UPV, ITACA, Spain): Medical Imaging

Funded by HoliBrain or external fundings 

• • Edern Le Bot (Ph.D, LaBRI): Deep learning

  • Sergio Ortega (Ph.D, UPV): Deep learning

  • Andrew Zamai (Ph.D, LaBRI): LLM

  • Eloi Navet (Ph.D, LaBRI): Deep learning

  • Trinh Dong (Postdoc, BPH): Statistic

Scientific Program

Objective A

The first scientific objective is to define the first multi-task and multi-level representation of brain anatomy. The division of brain analysis into many subtasks has led to a multitude of heterogeneous training databases with a large variety of partial and incompatible annotations (e.g., single-task or single-level). Recent efforts have been made to define harmonized brain annotation protocols. However, these consortiums only focus on one structure due to the complexity of the task and the impracticable time required for manual annotation (i.e., two full weeks to manually segmenting one brain at the structure level). To overcome this major issue, HoliBrain proposes first to define a holistic ontology of the whole brain anatomy by merging existing task- specific brain definitions in collaborations with experts. Second, HoliBrain proposes to build the first holistic training database by automatically fusing available partial annotations into a high-resolution multimodal dataset.

Objective B

The second scientific objective is to develop methods producing a complete analysis of cerebral anatomy from brain level to substructure level. DL has been studied intensively for segmentation of specific structures. However, the evolution towards a holistic brain segmentation raises several major challenges. First, the current DL architectures cannot accurately segment the whole brain at structure level due the limited training data (<50 samples) compared to the number of structures considered (≈100). Second, DL methods cannot process a monomodal (e.g., single contrast) 3D MRI at standard resolution (≈1mm3) due to limited GPU memory (typically 24GB). To move towards the proposed holistic paradigm comprising a large number of structures (≈400) and based on high- resolution (≈0.125mm3) multimodal MRI, major methodological breakthroughs must be proposed. To this end, HoliBrain proposes a collective artificial intelligence based on a large number of compact and collaborative DL networks.

Objective C

The third scientific objective is to produce new medical and neuroscientific knowledge. Major international initiatives have recently been launched to develop large-scale open access datasets. These big financial efforts (e.g., £250M for UKbiobank) were made against the promise of ground- breaking discoveries for patient care and society. However, the main problem is to use the existing databases at standard resolution (≈1mm3) with limited modalities to analyze substructures which are visible only at high resolution (≈0.125mm3) and with specific modalities. Thus, while previous works using specialized tools have demonstrated the great potential of analyzing these datasets, these studies used incomplete brain representation. The proposed holistic segmentation opens up new possibilities to simultaneously study brain evolution at different scales and to produce novel knowledge better characterizing the progression of pathologies. Therefore, HoliBrain proposes first to develop joint deep super-resolution and modality synthesis methods to reconstruct high- resolution data and to synthetize missing modalities in existing databases. Second, HoliBrain will perform bigdata analysis of the existing large-scale datasets with neurologists and neuroscientists to produce novel knowledge.

Objective D

The fourth scientific objective is to propose an efficient way of sharing the developed tools. Recently, open science and translational research have emerged as two fundamental issues in neuroscience and medical imaging. We think that it is very important to promote open science and translational research. Therefore, HoliBrain will integrate the developed flagship tools into an innovative, fully open access web platform to make them freely available and easily usable by the whole scientific community.

Publications:

1. Coupé, P., Mansencal, B., Morandat, F., Morell-Ortega, S., Villain, N., Manjón, J.V. and Planche, V., 2025. petBrain: A New Pipeline for Amyloid, Tau Tangles and Neurodegeneration Quantification Using PET and MRI. Alzheimer's Research & Therapy.

2. Bot, Edern Le, Rémi Giraud, Boris Mansencal, Thomas Tourdias, Josè V. Manjon, and Pierrick Coupé. "FLAIRBrainSeg: Fine-grained brain segmentation using FLAIR MRI only." In Medical Imaging with Deep Learning. 2025.

3. Coupé, Pierrick, Boris Mansencal, José V. Manjón, Patrice Péran, Wassilios G. Meissner, Thomas Tourdias, and Vincent Planche. "Lifespan tree of brain anatomy: diagnostic values for motor and cognitive neurodegenerative diseases." arXiv preprint arXiv:2502.09682 (2025).

4. Manjón, José V., Sergio Morell-Ortega, Marina Ruiz-Perez, Boris Mansencal, Edern Le Bot, Marien Gadea, Enrique Lanuza et al. "Ultra-high resolution multimodal MRI dense labelled holistic brain atlas." arXiv preprint arXiv:2501.16879(2025).

5. Longuefosse, Arthur, Edern Le Bot, Baudouin Denis De Senneville, Rémi Giraud, Boris Mansencal, Pierrick Coupé, Pascal Desbarats, and Fabien Baldacci. "Adapted nnU-Net: A Robust Baseline for Cross-Modality Synthesis and Medical Image Inpainting." In International Workshop on Simulation and Synthesis in Medical Imaging, pp. 24-33. Cham: Springer Nature Switzerland, 2024.

6. Ruiz-Perez, Marina, Sergio Morell-Ortega, Marien Gadea, Roberto Vivo-Hernando, Gregorio Rubio, Fernando Aparici, Mariam de la Iglesia-Vaya, Thomas Tourdias, Pierrick Coupé, and José V. Manjón. DeepThalamus: A novel deep learning method for automatic segmentation of brain thalamic nuclei from multimodal ultra-high resolution MRI. arXiv preprint arXiv:2401.07751 (2024).

7. Morell-Ortega, Sergio, Marina Ruiz-Perez, Marien Gadea, Roberto Vivo-Hernando, Gregorio Rubio, Fernando Aparici, Maria de La Iglesia-Vaya et al. "DeepCERES: A deep learning method for cerebellar lobule segmentation using ultra-high resolution multimodal MRI." NeuroImage (2025): 121063.

8. Perez-Caballero, Maria, Sergio Morell-Ortega, Marina Ruiz Perez, Pierrick Coupe, and Jose V. Manjon. Unsupervised Deep Learning Method for Bias Correction. In Medical Imaging with Deep Learning. 2024.