PMNet: A probability map based scaled network for breast cancer diagnosis

Area of Science:

• Oncology
• Medical Imaging
• Computer Science

Background:

• Breast cancer mortality is rising globally.
• Early detection is crucial for improving patient survival rates.
• Manual analysis of whole slide images for cancer detection is challenging for pathologists.

Purpose of the Study:

• To introduce PMNet, a pipeline for detecting invasive breast cancer regions in whole slide images.
• To classify whole slide images at the patch level into normal, benign, in situ, and invasive tumor categories.
• To improve the accuracy and efficiency of breast cancer detection in digital pathology.

Main Methods:

• Utilized scaled networks within the PMNet pipeline.
• Classified whole slide images into distinct categories (normal, benign, in situ, invasive).
• Developed a method for generating patch-level annotations for the TCGA breast cancer database.

Main Results:

• Achieved an f1-score of 88.9(±1.7)% on patch-level classification, outperforming the benchmark (81.2±1.3%).
• Obtained an average Dice coefficient of 69.8% on the BACH dataset, exceeding the benchmark (61.5%).
• Reached an average Dice coefficient of 82.7% on the dryad test dataset, surpassing the state-of-the-art (76%).

Conclusions:

• PMNet demonstrates superior performance in detecting invasive breast cancer in whole slide images.
• The developed pipeline offers a promising tool for enhancing early breast cancer detection.
• The proposed annotation generation method will support future deep learning research in breast cancer pathology.