Towards interactive AI-authoring with prototypical few-shot classifiers in histopathology

Area of Science:

• Digital pathology
• Medical artificial intelligence (AI)
• Computational pathology

Background:

• Histopathology analysis can greatly benefit from AI tools, with initial commercial products now available.
• Key challenges include the scarcity of annotated images and the need for models robust to data heterogeneity (domain generalization).

Purpose of the Study:

• To investigate the effectiveness of combining prototypical few-shot classification models with data augmentation for histopathology AI.
• To address the challenges of limited annotated data and domain generalization in AI-powered digital pathology.

Main Methods:

• Utilized prototypical few-shot classification models and domain-specific data augmentation on multi-center, multi-scanner datasets with two tumor entities.
• Compared nine state-of-the-art convolutional neural network (CNN) architectures for feature extraction, identifying EfficientNet B0 as optimal.
• Evaluated the impact of image prototypes from different scanners and the model's adaptability to a new tumor entity (urothelial carcinoma) with minimal annotations.

Main Results:

• Achieved ~90% classification accuracy on a multi-center, multi-scanner database, comparable to single-center performance, by using data from one scanner/site with augmentation.
• EfficientNet B0 demonstrated the best accuracy-inference time trade-off among tested CNNs.
• The few-shot model exhibited stable performance with minimal accuracy deviation (1.8%) when using cross-scanner prototypes.
• Adapted to classify urothelial carcinoma with 93.6% accuracy using only three annotations per subclass.

Conclusions:

• Prototypical few-shot classification is highly effective for AI in histopathology, requiring few annotations and enabling adaptation without extensive retraining.
• This approach facilitates the development of interactive AI authoring and guided annotation systems for non-technical users.