Weakly Supervised Deep Learning for Whole Slide Lung Cancer Image Analysis.

Area of Science:

• Digital pathology
• Computational oncology
• Machine learning in medicine

Background:

• Histopathology image analysis is crucial for cancer diagnosis, but computer-aided methods are underutilized.
• Whole slide image (WSI) classification faces challenges due to limited annotations, heterogeneous tumor patterns, and high computational costs.

Purpose of the Study:

• To develop a weakly supervised learning approach for efficient and effective classification of whole slide lung cancer images.
• To address the scarcity of annotations and computational challenges in WSI analysis.

Main Methods:

• A patch-based fully convolutional network (FCN) was used to extract discriminative features from image patches.
• Context-aware block selection and feature aggregation strategies were employed to create a holistic WSI descriptor.
• A random forest (RF) classifier was utilized for image-level prediction.

Main Results:

• The proposed method achieved a high accuracy of 97.3% on a large-scale lung cancer WSI dataset.
• It demonstrated superior performance compared to state-of-the-art approaches.
• The method also performed best on the public lung cancer WSIs dataset from The Cancer Genome Atlas (TCGA).

Conclusions:

• Weakly supervised learning, leveraging image-level labels and coarse annotations, is effective for WSI classification.
• The developed method offers a fast and feasible solution for computer-aided pathological diagnosis.
• This approach shows great potential to assist pathologists in near-future histology image diagnosis.