Code-free machine learning for classification of central nervous system histopathology images

Area of Science:

• Biomedical image analysis
• Artificial intelligence in pathology
• Computational pathology

Background:

• Machine learning (ML) offers transformative potential for biomedical image analysis, particularly in histopathology.
• Clinical adoption of ML is limited by the required informatics expertise.
• Code-free ML platforms aim to bridge this gap for broader accessibility.

Purpose of the Study:

• To evaluate the efficacy of two code-free ML platforms (Microsoft Custom Vision, Google AutoML) for classifying central nervous system histopathology images.
• To assess the feasibility of using these platforms for routine diagnostic tasks without extensive programming knowledge.

Main Methods:

• Histopathological images of central nervous system tissue samples were used to train Microsoft Custom Vision and Google AutoML.
• Classifications included glioma vs. brain metastasis, astrocytoma vs. astrocytosis, and 1p/19q co-deletion prediction.
• External validation and negative control experiments were employed to verify algorithmic accuracy.

Main Results:

• Both code-free ML platforms achieved high accuracy (80%–100%) in classifying histopathological images after training with 100–1000+ samples.
• Nontrivial classifications, including complex tumor differentiations and genetic marker predictions, were successfully performed.
• External validation and negative controls were critical for confirming the reliability of ML predictions.

Conclusions:

• Code-free machine learning platforms are suitable for accurate classification of diagnostic histopathology images.
• These platforms can facilitate the integration of AI into routine clinical pathology workflows.
• A proposed diagnostic workflow demonstrates the practical implementation of code-free ML for pathologists.