Back to Events

Seminar Series

January 17

On blood vessels and vascular networks

Bjoern Menze
University of Zurich

Part Of

2:00P
- 3:00P

Location

32-D451
Add to Calendar 2024-01-17 14:00:00 2024-01-17 15:00:00 America/New_York On blood vessels and vascular networks Vascular structures represent a part of the anatomy that attracts particular interest in biomedical imaging. Quantifying the structure of local vessels, as well as the global network, and making efforts to map blood flow and related functional physiological processes, is posing significant computational challenges in biomedical image analysis: Vessels need to be segmented in total to not miss on connections in the graph, foreground and background voxels are highly imbalanced, and training data for supervised learning approaches are tedious to annotate.I will present our work that improves image segmentation by synthesizing vascular structure for transfer learning, introduces new topological measures for segmenting vascular structures, and extracts network directly from image volumes using a transformer architecture that jointly predicts objects and their relations. I will focus on the extraction of whole brain vessel graphs of mice, capturing the full cerebrovascular network down to the capillary level, and I will comment on related benchmark dataset we made available at Neurips and MICCAI. 32-D451

December 13

Deep Image Analysis Based on Deformable Shapes and Its Application in Neuroimaging

Jian Wang
Harvard Medical School

Part Of

2:00P
- 3:00P

Location

32-D451
Add to Calendar 2023-12-13 14:00:00 2023-12-13 15:00:00 America/New_York Deep Image Analysis Based on Deformable Shapes and Its Application in Neuroimaging Deformable shapes are crucial for various image analysis tasks, e.g., image registration for real-time image-guided navigation systems of tumor removal surgery, image classification for neuro-degenerative diseases, and template-based image segmentation. Although recent advances in deep learning-based image analysis have achieved groundbreaking performance by providing a universal mechanism to extract image features in the context of texture, intensity, or simple geometry features, they fall short in capturing much more complex and detailed geometric information behind the image data. This greatly hinders the power of image analysis models when analyzing and quantifying geometric shapes are important. The modeling of deformable shapes presents significant challenges due to their high-dimensional and non-linear nature of data. Existing deep learning algorithms suffer from high computational cost of network training and inference, as well as severely decreased model performance caused by broken assumptions of image quality (i.e., missing data, corrupted signals, or occurrence of new objects). To address these challenges, I first developed deep neural networks to learn low-dimensional shape representations based on fine-grained deformations derived from image registration algorithms with much lower computational complexity in training. I then investigated a new paradigm of deep learning models that are capable of analyzing such learned shape representations to improve the current performance of image analysis tasks, including but not limited to population-based image studies. In order to enhance the robustness of shape-based deep networks, I developed geometric metamorphic learning algorithms to properly consider image conditions where missing data, or appearance changes occur. The algorithmic foundation of my research work could potentially impact a variety of real-world clinical applications, including but not limited to automated diagnosis for neurodegenerative diseases (i.e., Alzheimer’s) and real-time image-guided navigation systems for neurosurgery (i.e., brain tumor resection). 32-D451

November 15

Modality-Agnostic Representation Learning via Hierarchical Variational Auto-Encoders

Reuben Dorent
Harvard Medical School

Part Of

2:00P
- 3:00P

Location

32-D451
Add to Calendar 2023-11-15 14:00:00 2023-11-15 15:00:00 America/New_York Modality-Agnostic Representation Learning via Hierarchical Variational Auto-Encoders Learning pixel-level modality-agnostic representation of multi-modal imaging data is a challenging and open problem. In this work, Dr. Reuben Dorent will introduce MHVAE, a deep hierarchical variational auto-encoder (VAE) that generates modality-agnostic representations at the pixel level and allows for missing imaging modalities at training and testing time. Extending multi-modal VAEs with a hierarchical latent structure, a parametrization of the approximate posterior is introduced with a factorization similar to the true posterior, which can be expressed as a combination of unimodal variational posteriors. A simple optimization strategy is proposed to encourage learned representations to be modality-agnostic. Experiments on a database of intra-operative ultrasound (iUS) and Magnetic Resonance (MR) images demonstrate the effectiveness of the proposed approach at generating pixel-level representations that retain the content information while being similar for different sets of input modalities. 32-D451

October 25

Neural networks with Euclidean Symmetry for Learning from Physical Systems

Tess Smidt
MIT RLE

Part Of

2:00P
- 3:00P

Location

32-D451
Add to Calendar 2023-10-25 14:00:00 2023-10-25 15:00:00 America/New_York Neural networks with Euclidean Symmetry for Learning from Physical Systems Abstract:To use machine learning to tackle challenges in the chemical and biological sciences, we need methods built to handle the “datatypes” of physical systems: geometry and the geometric tensors. These are traditionally challenging datatypes to use for machine learning because coordinates and coordinate systems are sensitive to the symmetries of 3D space: 3Drotations, translations, and inversion.In this talk, I present a method that I have been developing with my colleagues for the past five years, Euclidean neural networks. These networks preserve Euclidean symmetry by construction, making them incapable of unphysical bias due to a change of coordinates. They eliminate the need for data augmentation -- the 500-fold increase in brute-force trainingnecessary for a model to learn 3D patterns in arbitrary orientations. This makes them extremely data-efficient; they result in more accurate models and require less training data to do so, which is ideal for modeling from scientific data that is expensive, difficult to acquire, or highly-varied.I describe how Euclidean neural networks work, demonstrate their effectiveness on a variety of real-world tasks, and introduce new capabilities my colleagues and I are developing with these methods. I also show how to efficiently and flexibly build equivariant models using ouropen-source PyTorch package e3nn (https://e3nn.org). 32-D451

September 20

Frozen Squirrels - Reconstructing 77 Mio Years of Evolution Towards the Human Brain

Georg Langs
Medical University of Vienna

Part Of

2:00P
- 3:00P

Location

32-D451
Add to Calendar 2023-09-20 14:00:00 2023-09-20 15:00:00 America/New_York Frozen Squirrels - Reconstructing 77 Mio Years of Evolution Towards the Human Brain This talk will discuss reconstructing the evolution of cortical geometry of the human brain and its link to function, behavior and ecology. A joint geometric representation of the cerebral cortices of ninety living species forms the basis of reconstructing ancestral shapes and tracing the evolutionary history of localised cortical expansions, modal segregation of brain function. We will discuss their association to behaviour and cognition, as individual cortical regions follow different sequences of area increase during evolutionary adaptations to dynamic socio-ecological niches. Anatomical correlates of this sequence of events are still observable in living species, and relate to their current behaviour and ecology. A decomposition of the evolutionary history of the shape of the human cortical surface into spatially and temporally conscribed components yields interpretable functional associations, and new evidence relating to the tethering hypothesis. 32-D451