Related Experiment Video
Updated: Aug 17, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Factorizer: A scalable interpretable approach to context modeling for medical image segmentation.
Pooya Ashtari1, Diana M Sima2, Lieven De Lathauwer3
1Department of Electrical Engineering (ESAT), STADIUS Center, KU Leuven, Leuven, Belgium; CREATIS (CNRS UMR5220 & INSERM U1294), Université Claude Bernard Lyon 1, Lyon, France.
This study introduces Factorizer models for medical image segmentation, achieving state-of-the-art results in brain tumor and stroke lesion segmentation. Factorizers offer improved accuracy, scalability, and interpretability compared to existing Convolutional Neural Networks and Transformers.
Area of Science:
- Medical image analysis
- Computer vision
- Machine learning
Background:
- Convolutional Neural Networks (CNNs) excel in medical image segmentation but struggle with global context due to convolution's locality.
- Transformers show promise in vision tasks but their quadratic attention complexity limits high-resolution context modeling.
- Existing methods face challenges in balancing global context capture with computational efficiency for medical image segmentation.
Purpose of the Study:
- To introduce a novel family of models, Factorizers, for end-to-end medical image segmentation.
- To address the limitations of CNNs and Transformers in capturing global context and computational scalability.
- To develop a linearly scalable approach for context modeling in medical image segmentation.
Main Methods:
- Proposed a U-shaped architecture integrating a differentiable Nonnegative Matrix Factorization (NMF) layer for linearly scalable context modeling.
- Utilized the shifted window technique in conjunction with NMF to effectively aggregate local information.
- Developed Factorizers, a novel model family leveraging low-rank matrix factorization for segmentation tasks.
Main Results:
- Factorizers achieved state-of-the-art performance on the BraTS (brain tumor) and ISLES'22 (stroke lesion) datasets.
- Demonstrated superior accuracy, scalability, and interpretability compared to CNNs and Transformers.
- NMF components provided enhanced interpretability over existing models.
- Achieved significant inference speed-up without substantial accuracy loss.
Conclusions:
- Factorizers represent a promising advancement in medical image segmentation, outperforming current state-of-the-art methods.
- The proposed NMF-based approach offers a scalable, interpretable, and efficient solution for complex segmentation tasks.
- Factorizers provide a unique advantage in terms of inference speed, making them highly practical for clinical applications.
