Extracting lung contour deformation features with deep learning for internal target motion tracking: a preliminary...

Area of Science:

• Medical Physics
• Radiotherapy
• Image-guided therapy

Background:

• Accurate tracking of internal target motion is crucial for effective radiotherapy.
• Current methods often require patient-specific data and extensive training.
• Non-invasive surrogates for motion estimation are highly desirable.

Purpose of the Study:

• To introduce lung contour deformation features (LCDFs) as a surrogate for estimating thoracic internal target motion.
• To evaluate the performance of LCDFs and a cascade ensemble model (CEM) in tracking motion.
• To assess the non-invasive nature and pre-treatment training requirements of the proposed methods.

Main Methods:

• LCDFs were extracted using an encoder-decoder deep learning model to match lung contours.
• A cascade ensemble model (CEM) estimated LCDFs from body images to track internal target motion.
• Performance was evaluated using motion data from 48 targets and compared to existing methods.

Main Results:

• LCDFs achieved a localization error of 2.6 ± 1.0 mm for internal targets.
• CEM demonstrated a localization error of 4.7 ± 0.9 mm with real-time performance of 256.9 ± 6.0 ms.
• Both methods showed minimal accuracy differences compared to patient-specific models without internal anatomy knowledge.

Conclusions:

• LCDFs and CEM effectively track target motion non-invasively.
• The proposed methods require no patient-specific training before treatment, simplifying clinical implementation.
• LCDFs and CEM hold significant potential for widespread use in radiotherapy and other image-guided interventions.