MedCLIP<T>

Models & Types · AiDotNet.VisionLanguage.Encoders

MedCLIP model using decoupled semantic matching for medical image-text alignment.

Inheritance: NeuralNetworkBase<T> → VisionLanguageModelBase<T> → MedCLIP<T>

Implements: IContrastiveVisionLanguageModel<T>

For Beginners

MedCLIP adapts CLIP for medical imaging by solving a key problem: medical datasets are small and images are not always paired with matching text. It uses decoupled semantic matching — any medical image can be paired with any text that shares the same medical concepts (like “pneumonia” or “chest X-ray”), allowing it to learn from unpaired data. Default values follow the original paper settings.

How It Works

MedCLIP (Wang et al., 2022) addresses limited medical data by decoupling image-text inputs: any image can be paired with any text sharing the same medical concepts (diagnosis, anatomy), using a semantic matching loss alongside contrastive learning.

References:

  • Paper: “MedCLIP: Contrastive Learning from Unpaired Medical Images and Text” (Wang et al., EMNLP 2022)

Example

var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.TwoDimensional,
    taskType: NeuralNetworkTaskType.ImageClassification,
    inputHeight: 224, inputWidth: 224, inputDepth: 3, outputSize: 512);
var trainModel = new MedCLIP<double>(architecture, new MedCLIPOptions());

Constructors

Constructor Summary
MedCLIP(NeuralNetworkArchitecture<T> architecture, MedCLIPOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer)
MedCLIP(NeuralNetworkArchitecture<T> architecture, string imageEncoderModelPath, MedCLIPOptions options)

Properties

Property Summary
T Temperature { get; } Gets the current learned temperature (the reciprocal of the clamped logit scale).
bool UsesBiasFreeReferenceProjections { get; } Gets whether both reference 512-dimensional projections are bias-free.
bool UsesTokenTypeEmbeddings { get; } Gets whether the ClinicalBERT embedding stack retains token-type embeddings.
int EmbeddingDimension { get; }
int MaxSequenceLength { get; }
int ProjectionDimension { get; }
int TextEmbeddingDimension { get; }
int TextTransformerBlockCount { get; } Gets the number of ClinicalBERT transformer blocks.
int VisionBottleneckBlockCount { get; } Gets the number of ResNet-50 bottleneck blocks.

Methods

Method Summary
Dictionary<string, T> ZeroShotClassify(Tensor<T> image, string[] labels)
ModelMetadata<T> GetModelMetadata()
T ComputeSemanticMatchingLoss(Tensor<T> imageEmbeddings, Tensor<T> textEmbeddings, Tensor<T> semanticScores) Computes MedCLIP’s bidirectional semantic matching objective for decoupled image/report batches.
T ComputeSimilarity(Tensor<T> image, string text)
Tensor<T> ComputeSemanticMatchingLossTensor(Tensor<T> imageEmbeddings, Tensor<T> textEmbeddings, Tensor<T> semanticScores) Computes the tape-connected MedCLIP semantic objective.
Tensor<T> EncodeImage(Tensor<T> image)
Tensor<T> EncodeText(string text)
Tensor<T>[] EncodeTexts(string[] texts)