New one-step model of breast tumor locating based on deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is the most prevalent cancer globally among women.
• Early detection is critical for effective treatment and reducing mortality.
• Automated tumor detection in breast ultrasound (US) images remains a challenge for computer-aided diagnosis (CAD).

Purpose of the Study:

• To develop a novel deep learning algorithm for automated breast tumor detection in US images.
• To improve the accuracy and efficiency of breast cancer diagnosis using artificial intelligence.

Main Methods:

• Proposed a new deep learning network, the 'One-step model', with one input and two outputs for segmentation and false-positive reduction.
• The model utilizes DenseNet for Base-net, RefineNet's decoder for Seg-net, and integrates Base-net and Seg-net layers into Cls-net.
• Employs an Anchor Box-based approach for lesion detection and a secondary network (Cls-net) for false-positive reduction.

Main Results:

• The One-step model achieved a 90.78% F1 score, outperforming the Single Shot MultiBox Detector (SSD) by 8.55%.
• Demonstrated computational efficiency and cost-effectiveness comparable to SSD.
• The model performed well on irregular and blurred ultrasound images.

Conclusions:

• The novel One-step model enhances location accuracy and reduces false targets for precise breast lesion detection.
• Achieved real-time tumor detection through a shared Base-net architecture.
• Confirmed the feasibility of deep learning for detecting breast lesions in US images.