BaT-SAM: Introducing Inductive Bias into SAM via a Mixture of Convolutional Experts for Brain Tumor Segmentation
We present BaT-SAM, a parameter-efficient adaptation framework that enables fully automatic brain tumor segmentation with the Segment Anything Model (SAM).
ProofPaper ↗
Key points
- Accurate brain tumor segmentation from magnetic resonance imaging (MRI) is essential for diagnosis and treatment planning; however, manual delineation is labor-intensive and time-consuming.
- BaT-SAM injects local inductive bias into SAM's frozen Vision Transformer (ViT) encoder via Conv-LoRA, where image features are processed by a sparsely gated mixture of multi-scale convolutional experts.
- To remove the need for user prompts, we further introduce a lightweight self-prompting box regressor that is jointly optimized with the adapter using scheduled sampling and predicts tumor locations directly from image embeddings.
- It further demonstrates strong cross-dataset generalization, achieving Dice scores of 0.702 on Figshare and 0.784 on the LGG MRI Segmentation dataset.
Sources (1)
- [1]BaT-SAM: Introducing Inductive Bias into SAM via a Mixture of Convolutional Experts for Brain Tumor SegmentationarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 04:09 AM
We present BaT-SAM, a parameter-efficient adaptation framework that enables fully automatic brain tumor segmentation with the Segment Anything Model (SAM).
Accurate brain tumor segmentation from magnetic resonance imaging (MRI) is essential for diagnosis and treatment planning; however, manual delineation is labor-intensive and time-consuming.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 10, 2026Toward System One Clinical Intelligence: Parameter-Efficient Fine-Tuning and Counterfactual Explainability of Decision Foundation Models
- Oct 10, 2026PGSB: Pretrained-Guided Shared Basis for LoRA Model Merging
- Oct 9, 2026Dream to Remember: Continual Steering of Generalist Robot Policies with World Models
- Oct 9, 2026Not All Dimensions Are Equal: Outlier-Aware Attention Linearization with Grouped Quadratic Kernels
- Oct 9, 2026Monarch-MoE: Structured Experts for Multimodal Continual Instruction Tuning
- Oct 8, 2026A Categorical Atlas of Parameter-Efficient Fine-Tuning: Cones, Orbits, and Gauges