Enhanced digital pathology image recognition via multi-attention mechanisms: the MACC-Net approach

Area of Science:

• Digital pathology
• Computational pathology
• Medical image analysis

Background:

• Manual interpretation of digital pathology images for cancer diagnosis is inefficient and subjective.
• Existing deep learning models struggle with recognizing osteosarcoma cell nuclei due to limitations in attention mechanisms and receptive fields.

Purpose of the Study:

• To develop a novel deep learning model, MACC-Net, to enhance the accuracy of osteosarcoma cell nucleus recognition in digital pathology images.
• To address the limitations of single-dimensional attention and fixed receptive fields in Convolutional Neural Networks (CNNs).

Main Methods:

• Introduction of MACC-Net, a multi-attention based deep learning model.
• Integration of channel, spatial, and pixel-level attention mechanisms.
• Enhancement of feature consistency and receptive field expansion for improved cell recognition.

Main Results:

• MACC-Net achieved a Dice Similarity Coefficient (DSC) of 0.847 in osteosarcoma cell nucleus recognition.
• The model demonstrated improved accuracy in edge recognition and differentiation of overlapping cells compared to traditional methods.

Conclusions:

• MACC-Net shows significant potential as a reliable auxiliary diagnostic tool for pathologists in digital pathology.
• The multi-attention approach effectively overcomes limitations of existing deep learning models for cancer cell nucleus recognition.