Unsupervised Learning Based on Multiple Descriptors for WSIs Diagnosis.

Area of Science:

• Computational pathology
• Medical image analysis
• Artificial intelligence in oncology

Background:

• Histopathological image analysis presents challenges due to cellular heterogeneity and limited information.
• Accurate automatic pathological diagnosis is crucial but difficult to achieve with existing methods.

Purpose of the Study:

• To develop an unsupervised deep learning model for whole-slide image diagnosis by fusing holistic and local appearance features.
• To enhance pathological analysis performance by overcoming limitations of heterogeneous data and intricate cellular structures.

Main Methods:

• Proposed an unsupervised deep learning model using stacked autoencoders to fuse multiple image descriptors (e.g., Histogram of Oriented Gradients, Local Binary Patterns) with original images.
• Extracted pre-trained latent vectors from autoencoders for classification.
• Developed a whole-slide image processing toolbox for patch extraction and processing.

Main Results:

• The model achieved high accuracies: 87.2% on ICIAR2018 and 94.6% on Dartmouth datasets.
• Demonstrated superior performance compared to state-of-the-art approaches in whole-slide image diagnosis.
• Visualization confirmed the model's ability to classify breast and lung cancer types similarly to pathologists.

Conclusions:

• The proposed unsupervised deep learning model effectively fuses heterogeneous features for enhanced whole-slide image diagnosis.
• The model's performance is robust across different datasets and does not rely on specific pre-trained classifier features.
• This approach offers a promising tool for accurate and efficient cancer classification in digital pathology.