NA-MIC Project Weeks
Back to Projects List
VolumeAXI - Volume Analysis, Explanability and Interpretability on CBCT
Key Investigators
- Jeanne Claret (University of Michigan, USA)
- Gaëlle Leroux (University of Michigan, USA)
- Lucia Cevidanes (University of Michigan, USA)
- Claudia Mattos (University of Michigan, USA)
- Juan Prieto (University of North Carolina, USA)
Project Description
Artificial Intelligences offer a wide range of applications. There is more and more research about its uses in the medical field and especially on patient images. The question of trustworthiness is on every mind, every time a prediction is given. For that reason, we propose to develop interpretable deep learning models for the automated classification of impacted maxillary canines and assessment of root resorption in adjacent teeth using Cone-Beam Computed Tomography (CBCT). Deep learning models based on Convolutional Neural Network (CNN) architectures were developed and evaluated for classifying impacted maxillary canine position and detecting root resorption. Gradient-weighted Class Activation Mapping (Grad-CAM) was integrated to generate visual explanations of the CNN predictions, enhancing interpretability and trustworthiness for clinical adoption.
Objective
We are using MONAI frameworks in this project.
- Data Preparation and Pre-processing
- Integrate the developed AI algorithms with clinical workflows.
- Enhancing Explainability and Interpretability in Medical Imaging by deploying a module in 3D Slicer
Approach and Plan
-
Data Preparation and Pre-processing
-
Model Development and Training: Explore and select appropriate neural network architectures (e.g., ResNet, SENets) for image classification and feature visualization.
-
Explainability and Visualization Techniques: Implement methods to make AI decisions transparent and understandable such as Grad-CAM.
-
Validation and Testing
-
Documentation and Training: Create comprehensive documentation and user guides explaining the functionality and benefits of the AI tools.
Progress and Next Steps
- Trained models with DenseNet architecture to classify the bucolingual position of the impacted maxillary canine and to classify the root resorption.
- Implementation of GRAD-CAM with MONAI for visualization
Next Steps:
- Find the best hyper-parameters for the given application to improve the results
- Deploy a module or an extension in 3D Slicer
Illustrations
Well predicted case
Class 1 predicted as a class 1. (right impacted canine)
The number of layers included change the precision of the focus on the tooth: