PatchSorter: a high throughput deep learning digital pathology tool for object labeling.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Accurate analysis of digital pathology images is crucial for diagnosis, prognosis, and therapy response prediction.
• Manual labeling of histological objects in large datasets is time-consuming and labor-intensive.
• Current methods lack efficiency for high-throughput analysis.

Purpose of the Study:

• To introduce PatchSorter, an open-source tool designed to accelerate the labeling of histological objects in digital pathology.
• To demonstrate the efficiency and accuracy of PatchSorter compared to traditional labeling methods.
• To facilitate high-throughput analysis of large digital pathology datasets.

Main Methods:

• Development of PatchSorter, an open-source labeling tool integrating deep learning with a web interface.
• Utilizing a dataset of over 100,000 histological objects for evaluation.
• Comparative analysis of labeling speed and accuracy between PatchSorter and unaided labeling.

Main Results:

• PatchSorter achieved a >7x improvement in labels per second compared to unaided labeling.
• The labeling accuracy remained minimally impacted, ensuring reliable data.
• Demonstrated feasibility for high-throughput labeling of extensive digital pathology datasets.

Conclusions:

• PatchSorter offers a significant advancement in the efficiency of digital pathology image analysis.
• The tool enables rapid and accurate labeling, overcoming previous bottlenecks in dataset annotation.
• Facilitates large-scale research in digital pathology for improved clinical insights.