MDFF-Net: A multi-dimensional feature fusion network for breast histopathology image classification

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Early breast cancer detection is critical for effective treatment.
• Deep learning-based computer-aided diagnosis (CAD) has improved medical diagnostics but faces limitations with complex or subtle pathological features.
• Current CAD systems can misdiagnose small lesions or deep cells due to insufficient feature extraction.

Purpose of the Study:

• To propose MDFF-Net, a novel Convolutional Neural Network (CNN)-based multidimensional feature fusion network.
• To enhance the accuracy and reduce misdiagnosis rates in pathological image analysis for breast cancer.
• To integrate 1D and 2D feature extraction for comprehensive image analysis, including cell morphology and differentiation.

Main Methods:

• Developed MDFF-Net, a CNN architecture incorporating 1D and 2D feature extraction networks and a feature fusion classification network.
• The 2D network utilizes multi-scale channel shuffling and channel attention modules.
• A 1D network, inspired by natural language processing, extracts detailed image information like cell morphology.

Main Results:

• MDFF-Net achieved 98.86% accuracy on the BreakHis dataset and 86.25% on the BACH dataset for breast cancer classification.
• The model demonstrated high efficacy on other cancer datasets, reaching 100% accuracy for colon and lung cancer classification.
• Experimental results indicate superior performance compared to classical classification models.

Conclusions:

• MDFF-Net effectively addresses limitations in current CAD systems by fusing multidimensional features.
• The proposed model shows significant potential for accurate and reliable cancer diagnosis across various pathological datasets.
• MDFF-Net offers a promising advancement in deep learning for medical image analysis and early disease detection.