Learning multi-frequency features in convolutional network for mammography classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer diagnosis relies heavily on mammography, but detecting small lesions with low contrast is challenging.
• Computer-aided diagnosis (CAD) systems offer decision support but face limitations with whole image classification.
• Lesion characteristics like size and contrast with surrounding tissue significantly impact diagnostic accuracy.

Purpose of the Study:

• To develop an advanced deep learning model for improved breast cancer detection in mammography.
• To enhance the accuracy of computer-aided diagnosis by highlighting salient features of potential tumors.
• To overcome challenges associated with classifying small lesions and low-contrast areas in mammographic images.

Main Methods:

• Introduction of the Multi Frequency Attention Network (MFA-Net), inspired by conventional machine learning.
• Decomposition of image features into low and high spatial frequency components using 2D Discrete Cosine Transform.
• Recalibration of discriminating features in separate frequency domains to focus on tumor presence and margins.

Main Results:

• The MFA-Net effectively mitigates the influence of irrelevant structures like pectoral muscle margins.
• Compared to traditional Convolutional Neural Networks (CNNs), MFA-Net demonstrates significant improvements in mammogram analysis.
• Achieved an Area Under the Curve (AUC) of 91.71% for malignant and benign classification on the INbreast dataset using transfer learning.

Conclusions:

• The proposed MFA-Net offers a robust approach for enhancing breast cancer detection in mammography.
• This method shows potential for improving the accuracy and reliability of computer-aided diagnosis systems.
• MFA-Net's ability to focus on critical lesion features represents a significant advancement in AI-driven medical imaging analysis.