|
Getting your Trinity Audio player ready...
|
Data is a cornerstone in this technology-driven era. It is a fundamental resource that empowers advancements across various domains. The more data available, the sharper and more accurate the outcomes it facilitates. However, in reality, obtaining complete data is only sometimes feasible or practical.
This limitation presents challenges that require innovative solutions. In light of this, Ulugbek Kamilov and Jiaming Liu, affiliated with the McKelvey School of Engineering at Washington University in St. Louis, embarked on a collaborative journey with researchers from the Lawrence Livermore National Laboratory.
They aim to develop a tool that maximises the utility of limited data, pushing the boundaries of what is achievable in data-driven technology.
This tool is known as DOLCE, an acronym that stands for “diffusion probabilistic limited-angle CT reconstruction.” DOLCE is a framework that harnesses the power of generative AI models to create multiple high-quality images even when the available data is severely limited.
Traditional generative AI models excel at creating realistic data, but they may need to catch up in terms of accuracy. DOLCE’s distinctive strength is that it marries the capabilities of AI with the ability to quantify its reconstruction uncertainty, ensuring reliability and trustworthiness.
“DOLCE allows to generate images that are realistic but consistent with the measured data, and it provides insights into the variance and uncertainty,” explained Kamilov.
Unlike traditional generative models, such as the large language models now familiar through chatbots, DOLCE consistently aligns its output with the actual data and the physical characteristics of the system it represents.
It also offers a variance map, which illustrates the range of possible image variants based on the available data. Kamilov said that this feature is pivotal in applications where trustworthiness and consistency are paramount.
It is important to note that while DOLCE is not designed for medical diagnostics, it offers a reliable perspective on what could be considered ground truth and the extent of variations possible in its generated images. This capacity is particularly vital in scenarios where collecting data from all angles is infeasible due to physical or temporal constraints.
“DOLCE epitomises this commitment to pushing the envelope, showcasing a computational imaging technology enabled by the latest generative modelling capabilities,” expressed Kamilov.
The impact of DOLCE extends far beyond the immediate applications in which it excels. It serves as an exemplar of generative modelling’s potential in the realm of computational imaging. This innovation fosters interdisciplinary collaboration, enabling experts from diverse fields to leverage DOLCE’s abilities in developing tailored solutions that push the boundaries of what can be achieved with advanced imaging technology.
The technology is a generative AI model at its core. It operates by generating multiple high-quality images when the available data is substantially limited. Generative AI models have proven their prowess in creating data that mimics reality.
However, one of their pitfalls is the potential lack of accuracy in the data they generate. In the DOLCE context, what sets it apart is the ability to bridge the gap between AI-generated data and data measured from the real world. This capacity ensures the consistency of its output with actual data, thus making it a highly reliable tool.
Furthermore, the range of DOLCE’s capabilities is demonstrated in two critical applications—airport security for scanned luggage and medical imaging of the human body. Through real-world limited-angle computed tomography (LACT) datasets, Liu effectively highlighted DOLCE’s prowess in generating high-quality images across a wide spectrum of scenarios. This adaptability positions DOLCE as a promising tool for elevating the quality of reconstructed LACT images.
DOLCE’s development underscores the relentless pursuit of extracting meaningful insights from even the most limited datasets. This collaboration is a testament to the quest for accuracy, reliability, and consistency, bridging the divide between data constraints and the limitless potential of AI-driven advancements.
