Semi-Supervised Deep Learning-Based Image Registration Method with Volume Penalty for Real-Time Breast Tumor Bed...

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Breast-conserving surgery necessitates radiotherapy to prevent recurrence.
• Accurate localization of the tumor bed for irradiation is challenging.
• Image registration can improve localization and reduce healthy tissue irradiation.

Purpose of the Study:

• To develop a novel deep learning-based nonrigid image registration method for breast tumor bed localization.
• To address data loss from tumor resection in radiotherapy planning.
• To enable real-time radiotherapy planning.

Main Methods:

• A modified U-Net architecture for deep learning-based nonrigid image registration.
• Multi-resolution processing to handle large deformations.
• A volume penalty incorporating tumor resection knowledge.

Main Results:

• The method achieved a mean target registration error below 6.5 mm.
• The relative volume ratio was close to zero, indicating accurate volume preservation.
• Registration time was under 1 second, enabling real-time application.
• Improvements were shown over classical and other learning-based methods.

Conclusions:

• The proposed method effectively localizes the breast tumor bed post-surgery.
• It enhances real-time radiotherapy planning by accurately registering pre- and post-operative scans.
• The approach reduces irradiation of surrounding healthy tissues, improving patient outcomes.