Related Experiment Video
Updated: Jul 3, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Optimization in Sparse 2D to Dense 3D Weakly Supervised Learning: Application to Multi-Label Segmentation of Large ex
Paul Hoareau1,2,3, Kuan Yi Wang1, Brandon Bujak4
1NeuroPoly Lab, Institute of Biomedical Engineering, Polytechnique Montreal, Montreal, QC, Canada.
Arxiv
|July 2, 2026
Summary
Weakly supervised learning for 3D spinal cord MRI segmentation requires distinct regularization strategies for 2D and 3D models. Human-centric enhancements can hinder, not help, 3D model performance.
Area of Science:
- Medical Imaging
- Machine Learning
- Neuroscience
Background:
- Fully supervised 3D segmentation of high-resolution ex vivo MRI is hindered by the high cost of volumetric annotation.
- Weakly supervised Sparse-to-Dense frameworks offer a solution but lack clear guidelines for cross-dimensional optimization.
- Divergent regularization needs for multi-class segmentation of ex vivo spinal cord MRI require analysis.
Purpose of the Study:
- To analyze divergent regularization needs for multi-class segmentation of high-resolution ex vivo spinal cord MRI.
- To evaluate the impact of human-centric preprocessing and spatial augmentation on 2D and 3D segmentation models.
- To investigate the transferability of optimization strategies across different dimensional architectures.
Main Methods:
- Utilized 9.4T MRI data from multiple sclerosis spinal cords with sparse annotations.
- Employed a 2D Teacher model trained on sparse slices to generate pseudo-labels for a 3D Student model.
- Systematically assessed the effects of human-centric preprocessing, spatial augmentation, and soft-label regularization.
Main Results:
- Identified critical divergence in training dynamics between 2D and 3D models.
- 2D Teacher benefited from strong spatial augmentation and soft-labeling, improving White Matter Lesion Dice scores by >11 points.
- Propagating these techniques to the 3D Student degraded performance; human-centric preprocessing reduced Gray Matter Lesion Dice scores by ~25 points.
Conclusions:
- Human-centric contrast enhancement can negatively impact machine learning model performance.
- 3D models trained on pseudo-labels require distinct, conservative regularization compared to 2D models.
- Optimization strategies must be tailored to the specific dimensional architecture and data characteristics.
