Visualizing histopathologic deep learning classification and anomaly detection using nonlinear feature space...

Area of Science:

• Digital pathology
• Computational biology
• Artificial intelligence in medicine

Background:

• Growing interest in AI and deep learning for histopathology computer vision.
• Current convolutional neural networks (CNNs) often rely on post-hoc analysis for classification.
• Lack of generalizable tools for visualizing deep learning inferences in histology.

Purpose of the Study:

• To develop a quantitative and transparent method for visualizing deep learning classifications in histopathology.
• To enable human review of deep learning decision-making processes.
• To create a generalizable classification and anomaly detection tool.

Main Methods:

• Utilizing t-distributed Stochastic Neighbor Embedding (t-SNE) for dimensionality reduction.
• Developing a workflow to visualize CNN organization of histomorphologic information.
• Discretizing class relationships on t-SNE plots for statistically-driven classifications.

Main Results:

• A quantitative and transparent approach to visualizing classification decisions before softmax compression.
• Super-imposing image regions onto t-SNE plots to render classifications based on distribution.
• Demonstrated automated, objective multi-class classifications with a priori defined cutoffs.

Conclusions:

• The novel approach provides intuitive outputs for human review and objective classification.
• This method serves as a generalizable tool for classification and anomaly detection, reducing reliance on post-hoc tuning.
• Aims to accelerate the adoption of CNNs in real-world histopathology applications.