UniverSeg<T>

Models & Types · AiDotNet.ComputerVision.Segmentation.Medical

UniverSeg: Universal Medical Image Segmentation via cross-attention.

Inheritance: NeuralNetworkBase<T> → SegmentationModelBase<T> → MedicalSegmentationBase<T> → UniverSeg<T>

For Beginners

Few-shot medical segmentation without fine-tuning. Cross-domain medical image segmentation.

Common use cases:

  • Few-shot medical segmentation without fine-tuning
  • Cross-domain medical image segmentation
  • Label-efficient medical AI
  • New task adaptation from a few examples

How It Works

Technical Details:

  • CrossBlock: each (query, support example) pair is convolved jointly and the query stream takes the

interaction averaged over the support set

  • A U-Net of CrossBlocks (64 features at four encoder and three decoder levels, LeakyReLU)
  • No fine-tuning needed for new segmentation tasks; set the support set with SetSupportSet
  • Trained on diverse medical segmentation datasets

Reference: Butoi et al., “UniverSeg: Universal Medical Image Segmentation”, ICCV 2023.

Example

// Create a UniverSeg model for few-shot medical segmentation
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.ThreeDimensional,
    taskType: NeuralNetworkTaskType.BinaryClassification,
    inputHeight: 128, inputWidth: 128, inputDepth: 1, outputSize: 1);
var model = new UniverSeg<double>(architecture);
// Every tunable is now set through the options object; these are the defaults:
var custom = new UniverSeg<double>(architecture,
    options: new UniverSegOptions { NumClasses = 1, DropRate = 0 });
            
// Or load a pre-trained ONNX model for cross-domain medical segmentation
var onnxModel = new UniverSeg<double>(architecture, "universeg.onnx");

Constructors

Constructor Summary
UniverSeg(NeuralNetworkArchitecture<T> architecture, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction, UniverSegOptions options) Initializes UniverSeg in native (trainable) mode.
UniverSeg(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, UniverSegOptions options) Initializes UniverSeg in ONNX (inference-only) mode.

Properties

Property Summary
bool Supports3D { get; } UniverSeg operates on 2D slices only.
bool SupportsFewShot { get; } UniverSeg is a few-shot model: its whole point is segmenting from a support set.

Methods

Method Summary
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
MedicalSegmentationResult<T> SegmentSlice(Tensor<T> slice) Segments a single 2D medical slice against the current support set.
MedicalSegmentationResult<T> SegmentVolume(Tensor<T> volume) Segments a volume by treating it as a single slice (UniverSeg is a 2D model).
ModelMetadata<T> GetModelMetadata() Collects metadata describing this model’s configuration.
Tensor<T> ForwardForTraining(Tensor<T> input) The training forward is the CrossBlock network.
void ClearSupportSet() Removes the support set; predictions then throw until a new one is set.
void SetSupportSet(Tensor<T> images, Tensor<T> labels) Sets the support set every prediction is conditioned on.