Improved FCOS for Detecting Breast Cancers.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is the leading cause of cancer-related death in women globally.
• Early detection significantly improves patient outcomes and survival rates.
• Accurate and automated detection algorithms are crucial for enhancing early diagnosis.

Purpose of the Study:

• To develop a fast, high-precision, and fully automated algorithm for breast cancer detection.
• To improve the early detection rate of breast cancer through advanced image analysis.
• To enhance the performance of existing object detection algorithms for breast lesion identification.

Main Methods:

• Comparison of anchor-based and anchor-free object detection algorithms for breast lesion detection.
• Implementation of the Fully Convolutional One-Stage Object Detection (FCOS) algorithm, an anchor-free method.
• Integration of a non-local technique to model long-range pixel dependencies for global context.
• Introduction of a novel deformable spatial attention (DSA) module to address variations in lesion shapes and sizes.

Main Results:

• The original FCOS algorithm achieved an average precision (AP) of 0.818 for benign and 0.888 for malignant lesions.
• The FCOS with a non-local module showed improved AP: 0.819 for benign and 0.894 for malignant lesions.
• Combining FCOS with the DSA module yielded the highest AP: 0.840 for benign and 0.899 for malignant lesions.

Conclusions:

• The proposed enhancements to the FCOS algorithm improve breast lesion localization and classification.
• The FCOS combined with the DSA module demonstrates significant potential for clinical application in ultrasound-guided breast cancer diagnosis.
• This advanced algorithm can provide valuable auxiliary diagnostic information for physicians, aiding in early and accurate breast cancer detection.