Synergizing Deep Learning-Enabled Preprocessing and Human-AI Integration for Efficient Automatic Ground Truth...

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Supervised deep learning in medical image interpretation is hindered by high costs and time for ground truth generation.
• Inconsistent image quality and overfitting in active learning sample selection are significant challenges.
• Whole-slide images (WSIs) offer rich data for automated histopathology, but inconsistencies persist.

Purpose of the Study:

• To address the limitations of ground truth generation in medical image interpretation using active learning.
• To mitigate overfitting in active learning by handling out-of-distribution samples caused by image inconsistencies.
• To improve the efficiency and accuracy of automated region-of-interest ground truth labeling on high-resolution WSIs.

Main Methods:

• Developed a novel deep learning-based preprocessing algorithm to normalize WSI data.
• Implemented an active learning strategy integrated with the preprocessing algorithm for sample selection.
• Quantitatively assessed and visually highlighted inconsistencies within the WSI dataset.

Main Results:

• The preprocessing algorithm effectively normalized unknown samples to the training set distribution, mitigating overfitting.
• Accepted 92% of automatically generated labels, expanding the labeled dataset by 845%.
• Achieved a 96% reduction in expert time compared to manual ground-truth labeling.

Conclusions:

• The proposed deep learning strategy significantly enhances automatic ground truth labeling for high-resolution WSIs.
• This approach effectively addresses image inconsistencies and reduces the cost and time of expert labeling.
• The method holds promise for advancing automated histopathology and medical image interpretation.