Enhancing Cone Beam CT Imaging with "One-Shot" Super-Resolution

Cone Beam Computed Tomography (CBCT) has become an indispensable tool in radiotherapy, offering quick and reliable imaging for high-precision treatments such as image-guided radiotherapy (IGRT). Its ability to provide real-time positional information about tumours and organs makes it invaluable for adaptive radiation therapy (ART). However, CBCT is plagued by low image quality compared to treatment planning CT (TPCT), limiting its utility in treatment planning and dose distribution calculations. Efforts to address these shortcomings have included deep learning-based image enhancement methods such as CycleGAN. Yet, these approaches typically require large amounts of training data, posing a significant challenge in medical imaging. A newly proposed "One-Shot" Super-Resolution (OSSR) method provides a promising alternative by substantially improving CBCT image quality with minimal data requirements.

The Challenges of CBCT Imaging

CBCT’s imaging principle results in several intrinsic limitations. Scattered rays and reconstruction artefacts lead to low contrast and suboptimal visual clarity, complicating the accurate identification of tumour boundaries and surrounding tissues. These shortcomings hinder the accurate calculation of dose distributions directly from CBCT images, a crucial step in adaptive treatment planning. Although combining CBCT with TPCT via non-rigid transformation offers a partial solution, it is often insufficient in areas with significant positional deviations, such as soft tissues or cavities in the intestinal tract.

Modern deep learning methods, such as U-Net and CycleGAN, have significantly advanced CBCT image enhancement. U-Net has been employed to improve CBCT images by training on datasets comprising pairs of CBCT and TPCT images. Similarly, CycleGAN enables domain transformation between CBCT and TPCT, allowing the generation of synthetic high-quality images. However, both methods rely heavily on large datasets, often requiring hundreds of image pairs from multiple patients. This dependency on extensive data collection is a substantial limitation in medical contexts, where data availability is constrained by ethical, logistical and privacy considerations.

The Innovation of OSSR

The "One-Shot" Super-Resolution (OSSR) method builds on the Zero-Shot Super-Resolution (ZSSR) technique, a deep learning-based approach that leverages the internal information of a single image for enhancement. Unlike traditional methods, OSSR requires only one pair of CBCT and TPCT images for training, making it particularly advantageous in medical settings where obtaining large datasets is impractical. By employing SimpleNet, a lightweight convolutional neural network, OSSR generates high-resolution CBCT images tailored to the specific characteristics of each image pair.

The OSSR process involves pairing each CBCT image with its corresponding TPCT image and applying data augmentation techniques, such as rotations, flips and cropping, to enhance training robustness. SimpleNet then processes these pairs to minimise the mean squared error (MSE) between the low-resolution CBCT and the high-resolution TPCT, resulting in a model optimised for image enhancement. The method's adaptability and efficiency enable it to address specific imaging challenges, such as the low contrast and artefacts inherent to CBCT, without requiring extensive pre-trained datasets.

Performance and Comparative Analysis

Quantitative evaluations of OSSR demonstrate its significant impact on CBCT image quality. Metrics such as Root Mean Squared Error (RMSE), Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) revealed marked improvements over unprocessed CBCT images and comparable or superior performance to established methods like CycleGAN. OSSR achieved an average reduction of 0.86 times in RMSE, reflecting fewer deviations between enhanced CBCT and TPCT images. PSNR values increased by 1.05 times, indicating reduced noise, while SSIM values improved by 1.03 times, signifying enhanced structural similarity as perceived by the human eye.

OSSR also outperformed traditional methods in terms of Normalised Mutual Information (NMI), a key metric for assessing positional accuracy. NMI values increased by an average of 1.31 times, surpassing both Total Variation Denoising (TVD) and CycleGAN. This improvement is particularly evident in anatomically complex areas, such as the rectum and bladder, where accurate positioning is critical. The OSSR method effectively suppresses artefacts, enhances tissue contours and improves overall image clarity, making it a valuable tool for high-precision radiotherapy.

A notable advantage of OSSR is its minimal data requirement. While CycleGAN typically requires datasets comprising hundreds of image pairs from multiple patients, OSSR achieves comparable results with just one pair of CBCT and TPCT images. This efficiency reduces the time and resources needed for data collection and model training, addressing one of the primary limitations of traditional deep-learning methods in medical imaging.

The "One-Shot" Super-Resolution (OSSR) method represents a significant advancement in CBCT imaging, offering a practical and efficient solution to longstanding challenges in adaptive radiotherapy. By leveraging minimal data, OSSR achieves notable improvements in image quality, matching or surpassing the performance of established methods such as CycleGAN. Its adaptability and efficiency make it particularly well-suited for clinical settings where data availability is limited.

OSSR’s ability to enhance CBCT image quality while preserving positional accuracy has important implications for radiotherapy. Improved CBCT images enable more accurate tumour localisation, better dose distribution calculations and enhanced treatment planning, ultimately improving patient outcomes. OSSR’s potential applications could extend to other anatomical regions, such as the head, neck and lungs, further broadening its impact in medical imaging. This innovative approach represents a crucial step towards more accurate and accessible radiotherapy solutions.

Source: Journal of Imaging Informatics in Medicine

Image Credit: iStock