by Christos Evangelou, MSc, PhD – Medical Writer and Editor
In a recent study, researchers conducted a virtual clinical trial to assess the potential of high-frequency computed tomography (CT) reconstruction to enhance imaging of liver metastases. This novel approach demonstrated superior preservation of crucial tumor features in low-noise conditions compared to standard reconstruction methods.1
Although the benefits of CT reconstruction diminish in the presence of typical clinical noise levels, the study opens new avenues for optimizing CT imaging in cancer diagnosis and monitoring, particularly in scenarios where high resolution and low noise are achievable.
“There is still much work to be done before this research can directly impact the clinic, but we are optimistic about the potential benefits,” said Christopher Wiedeman, PhD candidate at the Rensselaer Polytechnic Institute and first author of the study. “Ultimately, we believe that our findings, along with others, will direct the future practice to reconstruct medical images for both AI and human readers, who can complement each other to make better clinical decisions.”
The report was published in the journal of Visual Computing and for Industry, Biomedicine, and Art.
Study Rationale
Colorectal cancer, the third most commonly diagnosed cancer in the United States, often spreads to the liver, forming metastases that are a primary cause of mortality. Accurate imaging of these liver metastases is crucial for diagnosis, treatment planning, and monitoring disease progression.1 CT scans are a frontline tool for this purpose, but the preservation of important tumor features in CT images depends heavily on the scanning protocol and reconstruction algorithm used.
“Our main inspiration for investigating high-frequency kernels has been the increasing use of AI for medical image analysis,” Wiedeman noted. “Historically, CT reconstruction is tuned with human readers in mind, often at the cost of information loss.”
The team hypothesized that using a high-frequency reconstruction kernel in CT imaging could better preserve important features of liver metastases, potentially improving the accuracy of AI-based analysis of these tumors.
“If we tuned the process to preserve scan information, then the resulting images may appear less appealing to humans, but a data-driven reader could learn to make more informed predictions,” Wiedeman said.
Methodology: A Novel Virtual Clinical Trial
Researchers at Rensselaer Polytechnic Institute (NY, USA), GE HealthCare (NY, USA), Memorial Sloan Kettering Cancer Center, and Queen’s University (ON, Canada) teamed up to test the potential of high-frequency CT reconstruction to enhance the imaging of liver metastases using an innovative virtual clinical trial pipeline.1 This approach allowed them to generate and analyze a large number of simulated liver metastases under various imaging conditions without the ethical and practical constraints of using real patient data.
The team created a diverse set of 10,000 simulated liver metastases using a fractal-based method. These virtual tumors varied in characteristics such as edge sharpness, internal heterogeneity, and fractal dimension (a measure of edge complexity).1 Real CT scans of liver regions from patients with stage 1–3 colon cancer (without liver metastases) were used as background images. These were denoised and segmented to provide realistic contexts for the simulated tumors.1
Wiedeman explained that this fractal generation approach was inspired by a method used to create realistic coastlines and other self-similar shapes.
“Essentially, we start with a simple polygon and add both randomness and complexity to the shape by creating and perturbing new vertices between existing ones,” he said. “The result is a highly jagged shape that looks like an island. After this, we smooth the positions of adjacent vertices to varying degrees to modulate the overall smoothness of the shape border.”
Wiedeman also described how this approach allowed them to build a large population of diverse shapes and test different scanning procedures for their ability to preserve these characteristics.
“Generative algorithms that instead learn from CT images themselves are limited in this aspect by the image quality, constrained by the imaging system.”
Using a tool called CatSim,2,3 the researchers simulated CT scans of the combined background images and virtual metastases. The simulation parameters mimicked those of a GE HealthCare Lightspeed VCT scanner.1 Each simulated scan was reconstructed twice, once using a standard kernel and once using a high-frequency kernel.
The team then trained deep neural networks to recover the original tumor characteristics from the reconstructed images, allowing them to assess how well each reconstruction method preserved these features.1
High-Frequency Reconstruction Preserves Edge Sharpness in Noiseless Conditions
In the absence of noise, the high-frequency reconstruction method significantly outperformed the standard method in preserving edge sharpness and fractal dimension of the simulated metastases. The average squared error for characterizing edge sharpness was 12.2% lower with high-frequency reconstruction, while for fractal dimension, it was 7.5% lower.1
“We were pleased to see that the AI could ‘read through’ the aliasing artifacts and better estimate certain features with the high-frequency kernel,”
said Wiedeman. He added that these findings suggest that in scenarios where high-resolution, low-noise CT imaging is possible, such as focused scans of small regions of interest, high-frequency reconstruction could potentially provide more detailed and accurate tumor characterization.
However, when typical clinical noise levels were simulated, the performance difference between high-frequency and standard reconstruction became statistically insignificant for all tumor characteristics examined. Furthermore, the introduction of noise significantly degraded the ability to recover tumor features for both reconstruction methods, highlighting the critical role of noise in CT imaging. The ability to characterize internal tumor heterogeneity was comparable between the two reconstruction methods in both noisy and noiseless conditions.
According to the authors, even though the noisier appearance of high-frequency reconstructions may be less appealing to human observers, they could provide richer data for AI-based image analysis tools, potentially improving the accuracy of computer-aided diagnosis and prognosis.
Moreover, the study’s virtual clinical trial approach provides a framework for optimizing CT imaging protocols for specific clinical tasks, potentially leading to more tailored and effective imaging strategies in oncology.
“We envision that virtual clinical trials and simulation methods will become increasingly powerful in optimizing imaging procedures for tasks such as assessing colorectal liver metastases,” Wiedeman said.
Limitations and Future Work
Wiedeman acknowledged that, although the high-frequency kernel preserves more information, this information can be obfuscated by the increased noise associated with such a kernel.
“It is crucial to couple this change in the kernel with optimized scan techniques when studying the traits of these metastases,” he said.
Commenting on future work, Wiedeman highlighted that further clinical validation of this approach is needed.
“It is crucial to confirm whether a high-frequency kernel can better preserve some feature information associated with synthetic metastases for an AI reader in a real, high-quality retrospective cohort before any real clinical claims can be made,” he said.
He also noted that future work is needed to improve the clinical realism of the simulated liver metastases and to address the effects of image noise, adding that these could be overcome by combining data-driven and knowledge-based generative techniques and researching denoising algorithms.
The study received financial support from the NIH/NCI (grant no. R01CA233888) and the National Science Foundation Graduate Research Fellowship (grant no. DGE2147721).
References
- Wiedeman C, Lorraine P, Wang G, et al. Simulated deep CT characterization of liver metastases with high-resolution filtered back projection reconstruction. Vis Comput Ind Biomed Art. 2024;7(1):13. Published 2024 Jun 11. doi:10.1186/s42492-024-00161-y
- De Man B, Basu S, Chandra N, Dunham B, Edic P, Iatrou M et al (2007) CatSim: a new computer assisted tomography simulation environment. In: Proceedings of the SPIE 6510, medical imaging 2007: physics of medical imaging, SPIE, San Diego, 21 March 2007. https://doi.org/10.1117/12.710713
- Wu M, FitzGerald P, Zhang J, et al. XCIST-an open access x-ray/CT simulation toolkit. Phys Med Biol. 2022;67(19):10.1088/1361-6560/ac9174. Published 2022 Sep 28. doi:10.1088/1361-6560/ac9174
