FASNet: Feature alignment-based method with digital pathology images in assisted diagnosis medical system.

Area of Science:

• Digital pathology
• Medical imaging analysis
• Computational pathology

Background:

• Medical images, particularly digital pathology, are crucial for clinical diagnosis and tumor identification.
• Deep learning models for pathology image segmentation require extensive annotated data, which is costly and difficult to obtain.
• Current models struggle with domain shifts, leading to boundary prediction errors in new datasets.

Purpose of the Study:

• To develop a novel feature alignment-based detail recognition strategy for pathology image segmentation (FASNet).
• To improve the accuracy and robustness of cell nucleus identification in digital pathology images.
• To address the challenges of limited annotated data and domain shift in deep learning models.

Main Methods:

• Proposed FASNet, comprising a preprocessing model and a UNW segmentation network.
• Integrated semantics-aware normalization and whitening modules into the UNW network's encoder and decoder.
• Achieved feature compactness within classes and separation between classes for enhanced detail recognition.

Main Results:

• FASNet achieved a Dice Similarity Coefficient (DSC) of 0.844.
• Demonstrated robust performance on test data with distributions different from the training data.
• Successfully identified feature details for effective differentiation of tissue classes.

Conclusions:

• FASNet offers an efficient strategy for pathology image segmentation, improving detail recognition.
• The method enhances the ability to differentiate between various tissue classes.
• FASNet shows promise for reliable computer-aided diagnosis in digital pathology, even with domain variations.