The Utility of Unsupervised Machine Learning in Anatomic Pathology

Area of Science:

• Computational pathology
• Machine learning in medicine

Background:

• Supervised machine learning in anatomic pathology requires large annotated datasets, which are difficult and time-consuming to create.
• The majority of pathology data remains unlabeled, posing a significant challenge for developing accurate predictive models.

Purpose of the Study:

• To introduce unsupervised learning concepts.
• To illustrate how unsupervised methods like clustering, GANs, and autoencoders can address the lack of annotated data in anatomic pathology.

Main Methods:

• Literature review of unsupervised learning techniques.
• Examples of clustering, generative adversarial networks (GANs), and autoencoders applied to pathology data.

Main Results:

• Clustering facilitates semisupervised learning by propagating labels from annotated to unlabeled data.
• Generative adversarial networks (GANs) can generate synthetic data and perform color normalization.
• Autoencoders enable unsupervised pretraining on large unlabeled datasets, transferring learned features to classifiers trained on smaller labeled subsets.

Conclusions:

• Unsupervised machine learning techniques (clustering, GANs, autoencoders) offer solutions for the annotated data scarcity in pathology.
• These methods, individually or combined, can improve the development of supervised learning models in anatomic pathology.