FabNet: A Features Agglomeration-Based Convolutional Neural Network for Multiscale Breast Cancer Histopathology...

Area of Science:

• Digital pathology
• Computational imaging
• Machine learning in oncology

Background:

• Histopathology image diagnosis relies on expert experience, leading to inter-observer variability.
• Automated systems using deep learning can enhance diagnostic accuracy and reduce analysis time.
• Current models struggle with fine-to-coarse feature extraction in multi-scale histopathology images.

Purpose of the Study:

• To propose FabNet, a novel deep learning model for accurate histopathology image analysis.
• To improve the identification of malignant tumors by learning hierarchical features.
• To enhance diagnostic impartiality and reduce inter-operator variability in cancer diagnosis.

Main Methods:

• Developed FabNet, an accretive network architecture for multi-scale feature learning.
• Integrated deep and close feature combination across network layers.
• Utilized iterative and hierarchical feature map aggregation for classification.
• Trained and validated the model on breast and colon cancer histopathology datasets.

Main Results:

• FabNet achieved significant classification accuracy in identifying malignant tumors.
• The model demonstrated superior performance compared to state-of-the-art methods.
• Achieved high accuracy, F1 score, precision, and sensitivity with fewer parameters.
• Successfully identified malignant tumors from both whole images and image patches.

Conclusions:

• FabNet offers an efficient and accurate automated solution for histopathology image diagnosis.
• The model's hierarchical feature learning approach enhances diagnostic performance.
• FabNet has the potential to improve cancer detection and reduce diagnostic discrepancies.