Neural Scatter Removal for Image-Guided Radiation Therapy

May 2025 — August 2025

A U-Net that removes megavoltage scatter from treatment images to support image-guided radiation therapy.

PythonPyTorchU-NetOpenCVMedical Imaging

Neural Scatter Removal for Image-Guided Radiation Therapy

I designed a U-Net convolutional neural network to remove megavoltage scatter from kilovoltage images captured during cancer radiation therapy, producing clearer images for image-guided treatment.

The model improved validation-image PSNR by 46%, SSIM by 27%, and MAE by 73%, while maintaining strong performance on previously unseen rotational-scan data.

Highlights

Improved validation-image PSNR by 46%, SSIM by 27%, and MAE by 73%.
Generalized to previously unseen rotational-scan data.

Gallery

Comparison of kilovoltage, megavoltage-plus-kilovoltage, and synthetic radiation images — The synthetic output reduces scatter present in the combined MV + kV treatment image.

Intensity comparison between combined treatment images and synthetic kilovoltage output — Intensity comparison between the input image and synthetic kV-only result.

Line-profile analysis comparing the U-Net output with reference imaging data.