Summary:
Researchers in the Department of Radiological Sciences have developed a machine learning algorithm that processes multispectral photon counting CT data for accurate medical imaging.
Background:
Photon counting computed tomography (PCCT) is a tremendous engineering advancement, enabling high resolution spectral imaging with myriad applications. As opposed to traditional CT imaging methods which use an energy integrating detector, PCCT uses detectors that count individual photons. The photon counting detectors can discriminate photons based on their energy levels, potentially enabling superior tissue characterization since different tissues absorb and scatter X-rays differently depending on the composition of the tissue. This system creates two major benefits; 1) significantly improved spatial resolutions of CT images, leading to better diagnostic decision making and 2) lower doses of radiation for patients. Despite these benefits, the full utility of PCCT has yet to be realized due to the still highly manual nature of image analysis. A vast number of multi-energy images are available from PCCT, but due to practical time constraints, radiologists and other interpreters only examine a subset of these images to reach diagnostic conclusions. There is a need for automated, accurate methods to screen through the full spectrum of PCCT-derived datasets to provide the greatest utility to patients.
Innovation:
Researchers led by Professor Matthew Brown have developed a software algorithm that applies deep neural networks to process medical images at high throughput and accuracy. The neural networks are trained on full spectrum PCCT data. The system employs a novel training and optimization process to optimize the algorithm for identification of the most accurate neural network and repeats this process until the best algorithm is identified. This system could be easily modified for different types of tissues, imaging planes, and diagnostic indications. By incorporating this software with pre-existing clinical hardware, it will be possible to greatly improve radiological workflows and improve patient outcomes with more detailed, accurate diagnostics.
Potential Applications:
• PCCT image analysis
• Automated radiology
• Cancer prognostics
• Digital pathology
Advantages:
• Efficient, automated PCCT image processing
• Utilizes entire dataset generated by instrumentation
• Configurable to different tissue and disease types
Development-To-Date:
The base software has been developed and validated on non-PCCT image datasets.
Related Papers:
Choi Y, Wahi-Anwar MW, Brown MS (2023) SimpleMind: An open-source software environment that adds thinking to deep neural networks. PLoS ONE 18(4): e0283587. https://doi.org/10.1371/journal.pone.0283587
Reference:
UCLA Case No. 2024-058
Lead Inventor:
Matthew Brown