Factorizer

Factorizer applies Nonnegative Matrix Factorization (NMF) within a U-shaped architecture to perform end-to-end medical image segmentation that models global context and produces interpretable components for tasks such as brain tumor and stroke lesion segmentation.

Key Features:

• End-to-End Segmentation Model: Integrates NMF as a differentiable layer inside a U-shaped architecture to enable end-to-end training and inference.
• Linearly Scalable Context Modeling: Formulates NMF in a linearly scalable manner to avoid the quadratic complexity of attention mechanisms while capturing global context.
• Shifted Window Technique: Employs a shifted window technique alongside NMF to aggregate local information and enhance recognition of fine structures.
• Interpretability via NMF Components: Produces meaningful NMF components that provide interpretable factors for segmented medical images compared with standard CNNs and Transformers.
• Performance and Efficiency: Achieves state-of-the-art results on benchmark datasets such as BraTS (brain tumor segmentation) and ISLES'22 (stroke lesion segmentation) and demonstrates inference speed improvements reported by ablation studies.

Scientific Applications:

• Clinical medical image segmentation: Accurate segmentation of anatomical structures and pathological regions in clinical imaging data.
• Brain tumor segmentation (BraTS): Applied to glioma and brain tumor delineation tasks evaluated on the BraTS benchmark.
• Stroke lesion segmentation (ISLES'22): Applied to ischemic stroke lesion delineation tasks evaluated on the ISLES'22 benchmark.
• Interpretability for clinicians: Uses NMF-derived components to aid interpretation and analysis of segmentation results by clinical users.

Methodology:

Integrates low-rank Nonnegative Matrix Factorization (NMF) as a differentiable layer within a U-shaped architecture, formulates NMF for linear scalability to replace quadratic attention complexity, employs a shifted window technique to aggregate local information, and trains the model end-to-end.