NA-MIC Project Weeks

Back to Projects List

ShapeAXI Shape Analysis Exploration and Interpretability

Key Investigators

• Juan Prieto (University of North Carolina, USA)
• Nathan Hutin (University of Michigan)
• Luc Anchling (University of Michigan, France)
• Lucia Cevidanes (University of Michigan, USA)
• Selene Barone (University of Catanzaro, Italy)
• Jonas Bianchi (University of Pacifique, USA)
• Marcela Gurgel (University of Michigan, USA)
• Najla Al Turkestani (University of Michigan, USA)
• Felicia Miranda (University of Sao Paulo, Brezil)
• Denise Curado (University of Michigan, USA)
• Kinjal Mavani (University of Michigan, USA)
• Kinjal Mavani (University of Michigan, USA)
• Margaret Eason (University of Michigan, USA)
• Aron Aliage del Castilo (University of Michigan, USA)

Presenter location: In-person

Project Description

ShapeAXI is an innovative project that focuses on advancing the field of shape analysis, exploration, and interpretability through the application of artificial intelligence (AI) techniques. The project aims to develop novel algorithms and tools that can effectively analyze and interpret complex shapes, enabling deeper insights and understanding in various domains such as computer vision, computer graphics, and biomedical imaging. By harnessing the power of AI, ShapeAXI aims to revolutionize shape analysis by automating the process of shape exploration, identifying patterns, and extracting meaningful information. The project strives to enhance interpretability, enabling users to comprehend and interpret the underlying structure and characteristics of shapes with greater clarity. ShapeAXI holds the promise of unlocking new possibilities in shape-related research and applications, ultimately leading to advancements in fields such as object recognition, shape synthesis, and shape-based decision-making systems.

Objective

1. Develop advanced shape analysis algorithms: Design and develop cutting-edge algorithms that can efficiently analyze and process complex shapes, encompassing both 2D and 3D domains. These algorithms will leverage AI techniques such as deep learning, computer vision, and geometric modeling to provide accurate and robust shape analysis capabilities.

2. Enable shape exploration and discovery: Create tools and techniques that allow users to explore and navigate through shape spaces effectively. By leveraging AI-driven approaches, the project will enable users to discover shape patterns, similarities, and differences, facilitating insights into shape characteristics and structures.

3. Enhance shape interpretability: Develop methods to enhance the interpretability of shape analysis results, enabling users to understand and interpret the underlying meaning and significance of shape features. This includes visual explanations, feature attribution techniques, and intuitive representations to facilitate human comprehension of shape analysis outcomes.

4. Foster cross-domain applicability: Ensure the developed shape analysis techniques and tools are applicable across various domains, such as computer vision, computer graphics, biomedical imaging, and manufacturing. The project will focus on creating adaptable and versatile solutions that can be effectively utilized in different application areas.

5. Promote open-source collaboration: Foster a collaborative and open-source approach to encourage knowledge sharing and community involvement. The project will aim to release relevant software libraries, datasets, and benchmarks, allowing researchers and practitioners to build upon the developed tools and algorithms and advance the field collectively.

6. Validate and benchmark performance: Conduct extensive validation experiments and comparative studies to assess the performance and efficacy of the developed algorithms and tools. This includes benchmarking against existing methods and datasets, ensuring the reliability and generalizability of the proposed shape analysis techniques.

Approach and Plan

1. Develop algorithms for shape analysis, incorporating deep learning techniques, computer vision, and geometric modeling.
2. Create a user-friendly 3D Slicer extension with a software interfaces for creating the explainability maps on shapes after classification.
3. Enhance interpretability through techniques like visual explanations and feature attribution methods.
4. Validate the developed algorithms and tools through extensive experiments and benchmarking against existing methods.
5. Foster open-source collaboration by releasing software libraries, datasets, and benchmarks.
6. Document methodologies and findings for knowledge sharing and prepare technical papers and presentations.

Progress and Next Steps

1. Develop the algorithms for shape analysis and classification.
2. Create heat maps using GradCAM and propagate them to the shapes

Illustrations

Results

Inputs

surf	class
/path/to/model.vtk	0
/path/to/model2.vtk	1
/path/to/model3.vtk	2

Automated training - testing

Explainability

Background and References

1. https://github.com/DCBIA-OrthoLab/Fly-by-CNN.git
2. Selvaraju, Ramprasaath R., Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and Dhruv Batra. “Grad-CAM: Why did you say that?.” arXiv preprint arXiv:1611.07450 (2016).