NA-MIC Project Weeks

Back to Projects List

ALI: Automated Landmark Identification update

Key Investigators

• Jeanne Claret (University of Michigan, USA)
• Gaëlle Leroux (University of Michigan, USA)
• Lucia Cevidanes (UoM, USA)
• Juan Prieto ( University of North Carolina, USA)

Presenter location: Online

Project Description

This is an update to a Slicer module developed during project week #37. The approach reformulates anatomical landmark detection as a classification problem through a virtual agent placed inside a 3D Cone-Beam Computed Tomography (CBCT) scan. This agent is trained to navigate in a multi-scale volumetric space to reach the estimated landmark position. The agent movements decision relies on a combination of Densely Connected Convolutional Networks (DCCN) and fully connected layers. Automated Landmark Identification (ALI) is a tool available in the extension SlicerAutomatedDentalTool. This module aims to automatically identify landmarks on different type of scans (CBCT, IOS). The current CBCT models trained include the Cranial Base, Upper and Lower Bones of the face, and the teeth (Left, Right, Upper, and Lower).

Objective

1. Retrain the different landmarks models with new and larger datasets annotated by clinicians, with particular focus on teeth on atypical or impacted position inside the bone.
2. Hyperparameters fine tuning for maintenance and improved accuracy on the previously available code.

Approach and Plan

1. Collected and preprocessed data
2. Model Architecture: Review and improve the existing model architecture, which utilizes Densely Connected Convolutional Networks (DCCN) and fully connected layers.
3. Hyperparameter Fine-Tuning: Conduct hyperparameter optimization to fine-tune the model for improved accuracy. This includes adjusting learning rates, batch sizes, and regularization techniques.
4. Training and Validation Dataset Split: Divide the dataset into training, validation, and test sets to ensure proper model evaluation. Model Training: Train the models for automated landmark identification using the restructured dataset. Employ techniques such as data augmentation to improve generalization. Validation Metrics: Use appropriate evaluation metrics such as maximum, mean and standard deviation fo errors compared with gold standard annotations.
5. Pull request of the updated models into the SlicerAutomatedDentalTool extension to enable improved automated landmark identification for CBCT scans.
6. User Interface: Enhance the user interface to make it user-friendly and intuitive for clinicians.
7. Testing and Quality Assurance: Thoroughly test the updated module to identify and resolve any bugs or issues. Ensure that the automated landmark identification module performs accurately and reliably on different types of scans.
8. Documentation and Training: Create comprehensive documentation for users and developers, including instructions on how to use the module effectively.

Progress and Next Steps

1. Collected and preprocessed data
2. Hyperparameter Fine-Tuning
3. Training Dataset Split: Divide the dataset into training, validation, and test sets to ensure proper model evaluation.
4. The evaluation metrics currently seems unreliable; so I am seeking guidance on how to improve.

Illustrations

Background and References

Link to the AutomatedDentalTool Github: https://github.com/Jeanneclre/SlicerAutomatedDentalTools/tree/main