SegMamba<T>
Models & Types · AiDotNet.ComputerVision.Segmentation.Medical
SegMamba: long-range sequential modeling Mamba for 3D medical image segmentation (Xing et al., 2024, arXiv:2401.13560).
Inheritance: NeuralNetworkBase<T> → SegmentationModelBase<T> → MedicalSegmentationBase<T> → SegMamba<T>
For Beginners
This model labels every voxel of a 3-D medical scan (e.g. a CT volume) with the organ/structure it belongs to. It “reads” the whole volume as a long sequence in several directions so it can relate far-apart regions cheaply.
How It Works
Architecture (paper-faithful). SegMamba is a 3D U-Net whose encoder replaces the usual self-attention / convolution stack with a Mamba state-space backbone:
The Mamba scan gives linear complexity in the number of voxels, which is what makes whole-volume 3D segmentation tractable where attention would be quadratic.
Reference: Xing et al., “SegMamba: Long-range Sequential Modeling Mamba For 3D Medical Image Segmentation”, arXiv:2401.13560, 2024.
Constructors
| Constructor | Summary |
|---|---|
SegMamba(NeuralNetworkArchitecture<T> architecture, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction, SegMambaOptions options) |
Initializes SegMamba in native (trainable) mode. |
SegMamba(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, SegMambaOptions options) |
Initializes SegMamba in ONNX (inference-only) mode. |
Properties
| Property | Summary |
|---|---|
bool Supports2D { get; } |
SegMamba is a purely volumetric (3D) model; single 2D slices are not supported. |
Methods
| Method | Summary |
|---|---|
Dictionary<string, Tensor<T>> GetNamedLayerActivations(Tensor<T> input) |
|
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) |
|
MedicalSegmentationResult<T> SegmentFewShot(Tensor<T> queryImage, Tensor<T> supportImages, Tensor<T> supportMasks) |
|
MedicalSegmentationResult<T> SegmentSlice(Tensor<T> slice) |
|
MedicalSegmentationResult<T> SegmentVolume(Tensor<T> volume) |
|
ModelMetadata<T> GetModelMetadata() |
|
Tensor<T> ForwardForTraining(Tensor<T> input) |