Real-time tumor localization with single x-ray projection at arbitrary gantry angles using a convolutional neural...

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Real-time tumor localization is critical for effective tumor tracking therapy.
• Existing single X-ray projection methods using convolutional neural networks (CNNs) are limited to fixed gantry angles.
• Rotational radiotherapy requires tumor localization techniques adaptable to arbitrary gantry angles.

Purpose of the Study:

• To develop and validate an improved CNN-based technique for real-time tumor localization from single X-ray projections at arbitrary gantry angles.
• To address the challenges of gantry rotation in tumor motion tracking for radiotherapy.
• To enhance the applicability of tumor localization frameworks for rotational therapy.

Main Methods:

• A novel CNN incorporating a binary region of interest (ROI) mask was developed to mitigate overfitting caused by gantry rotation.
• An angle-dependent fully connected layer (ADFCL) was utilized to map extracted features to tumor motion, accounting for varying gantry angles.
• The method was evaluated using X-ray projection data from 15 realistic patients, comparing results with a VGG network variant.

Main Results:

• The proposed CNN achieved an average tumor localization error under 1.8 mm (superior-inferior) and 1.0 mm (lateral) for patients with clearly visible tumors.
• For patients with less visible tumors, a feature point localization error of no more than 1.5 mm (superior-inferior) and 1.0 mm (lateral) was achieved.
• The technique demonstrated robust performance across diverse patient data and arbitrary gantry angles.

Conclusions:

• A novel CNN-based method for real-time tumor localization from single X-ray projections at arbitrary angles has been successfully developed and validated.
• This approach significantly expands the applicability of tumor localization to rotational radiotherapy.
• The technique offers precise and reliable tumor tracking, crucial for advancing radiation therapy precision.