A multi-level feature-fusion-based approach to breast histopathological image classification.

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Convolutional neural networks (CNNs) primarily utilize deep semantic features for image classification.
• Deep features enhance target classification but often neglect crucial shallow local features like texture and edge contours in histopathology images.
• This limitation reduces the accuracy of breast cancer detection from histopathology images.

Purpose of the Study:

• To develop an advanced multi-level feature fusion method for improved breast histopathology image classification.
• To address the loss of shallow local features in traditional CNN-based approaches.
• To enhance the accuracy and reliability of automated breast cancer diagnosis.

Main Methods:

• A novel multi-level feature fusion approach combining shallow and deep semantic features using attention mechanisms and convolutions.
• Implementation of a weighted cross-entropy loss function to mitigate misclassifications (false positives and false negatives).
• Utilizing spatial information correlation to refine patch-level diagnostic accuracy.

Main Results:

• The proposed method achieved a classification accuracy of 99.0% on experimental datasets.
• An Area Under the Curve (AUC) of 99.9% was obtained, indicating excellent diagnostic performance.
• The approach demonstrated superior performance compared to the baseline Inception-ResNet-v2 network.

Conclusions:

• The multi-level feature fusion method effectively integrates diverse feature levels for robust breast histopathology image classification.
• The proposed techniques significantly improve diagnostic accuracy and reduce misjudgments in automated cancer detection.
• This approach holds promise for advancing computer-aided diagnosis in digital pathology.