Scale-Adaptive viable tumor burden estimation via histopathological microscopy image segmentation.

Area of Science:

• Digital pathology
• Computer vision
• Medical image analysis

Background:

• Accurate cancer segmentation in whole-slide images is vital for tumor burden estimation and cancer assessment.
• Challenges include vague boundaries and small, dissociated tumor regions, complicating segmentation.
• Multi-scale features are beneficial for addressing these complexities in vision tasks.

Purpose of the Study:

• To develop a structure-aware, scale-adaptive feature selection method for efficient and accurate cancer segmentation.
• To improve the representation of vague, non-rigid boundaries and enhance the segmentation of small tumor regions.
• To evaluate the proposed method against state-of-the-art techniques in liver and colorectal cancer segmentation.

Main Methods:

• A segmentation network with an encoder-decoder architecture was utilized.
• A scale-adaptive module was developed to select robust features for boundary representation.
• A structural similarity metric was introduced to improve tissue structure awareness and small region segmentation.
• Attention mechanisms and selective-kernel convolutions were incorporated for comparative analysis.

Main Results:

• The proposed structure-aware, scale-adaptive network achieved superior performance in liver cancer segmentation, surpassing top results from the PAIP 2019 challenge.
• For colorectal cancer segmentation, the scale-adaptive module improved the baseline network's performance or outperformed advanced attention mechanisms.
• The method demonstrated a favorable trade-off between efficiency and accuracy in cancer segmentation tasks.

Conclusions:

• The developed structure-aware, scale-adaptive network offers an effective solution for accurate and efficient cancer segmentation in whole-slide images.
• The scale-adaptive module is particularly valuable for handling challenges like vague boundaries and small tumor regions.
• This approach shows significant potential for improving diagnostic accuracy in digital pathology.