Weakly Supervised Learning using Attention gates for colon cancer histopathological image segmentation.

Area of Science:

• Artificial Intelligence
• Digital Pathology
• Computational Biology

Background:

• Deep learning methods have revolutionized various applications, including digital pathology for tumor diagnosis and prognosis.
• Classical machine learning methods struggle with Whole Slide Images (WSI) due to their large size, high resolution, and limited annotated samples, hindering generalization.
• Traditional methods exhibit poor generalization across different tasks and data types in histopathological image analysis.

Approach:

• This study explores deep learning models, specifically UNet and Att-UNet, for colon cancer WSI segmentation in sparsely annotated datasets.
• Novel enhanced Att-UNet models are introduced, optimizing skip connections and spatial attention gate placement for improved feature learning.
• A multi-step training strategy is proposed to address data scarcity, sparse annotations, and class imbalance in colon cancer datasets.

Key Points:

• Spatial attention gates enhance training by preventing irrelevant feature learning, leading to more robust segmentation.
• The Alter-AttUNet model offers a balance between accuracy and network efficiency, achieving 95.88% accuracy on the AiCOLO colon cancer dataset.
• Proposed methods outperform state-of-the-art approaches on both proprietary (AiCOLO) and public datasets (NCT-CRC-HE-100K, CRC-5000, Warwick).

Conclusions:

• The developed Alter-AttUNet model provides a robust and accurate solution for histopathological image segmentation in digital pathology.
• The multi-step training strategy effectively handles sparse annotations and class imbalance, crucial for real-world datasets.
• The findings demonstrate the potential of advanced deep learning techniques to improve the efficiency and accuracy of cancer diagnosis from histopathological images.