To this end, among the topics discussed at the Arena session on January 26th was whether deep learning AI algorithms could replace traditional radiomics analysis.
“AI is now officially a buzzword in our field, and it’s very exciting,” said moderator Dr. Tyler Bradshaw, assistant professor of radiology at the University of Wisconsin-Madison. “But about ten years ago there was another buzzword that promised a lot of the same thing. That buzzword was Radiomics.”
Radiomics, which refers to the extraction of minable data from medical images, has traditionally been performed using machine learning algorithms trained to extract specific image features for analysis. I came. According to RSNA, the technology is being applied to improve diagnosis, prognosis and clinical decision support for the delivery of precision medicine.
AI and radiomics experts discussed the promising capabilities of the two approaches at the “Arena Session” held on June 26th at SNMMI 2023.
Conversely, RSNA defines deep learning as a class of machine learning that can be “trained” to automatically detect these features, unlike radiomics, which requires manual extraction of features from input images. I’m here.
“The literature is full of examples showing that very high accuracy can be achieved using deep learning-based features rather than traditional hand-crafted radiomics-based features,” says Dr. Joita Dutta. said. Associate professor of biomedical engineering at the University of Massachusetts Amherst.
Specifically, she said, neural networks show the potential to automatically identify the parts of an image that are most relevant to the task of interest. This eliminates the need to segment images individually, she said. In other words, a deep learning approach could reduce the burden on doctors, she said.
“So for me, it’s a no-brainer,” Dutta said.
Dr. Abhinav Ja of Washington University in St. Louis also argued in favor of deep learning algorithms over traditional radiomics. AI algorithms may be more reproducible and reliable, he said, and can learn from large datasets to identify “hidden” features within tumors that traditional radiomics techniques cannot. Stated.
Furthermore, the medical image datasets used in radiomics can be heterogeneous due to the use of different scanners and protocols for image acquisition, whereas deep learning is based on the “universal approximation theorem”. , he said, given enough data to an AI algorithm, it would be able to mimic most functions.
“With enough data, deep learning could potentially model heterogeneity due to variations in scanners and image processing protocols,” Jha said.
The future of radiomics
Importantly, deep learning AI algorithms have so far not outperformed traditional radiomics methods, said Dr. Irene Bubat, arguing that the future of radiomics remains bright. .
Buvato, Head of the In Vivo Molecular Imaging Laboratory at the Service Hospitalier Frédéric Joliot PET Center in Orsay, France, spoke about the competition at the 2022 Conference on Medical Image Computing and Computer-Assisted Interventions. As a challenge, participants developed different models to predict recurrence-free survival in head and neck cancer patients from F-18 FDG-PET/CT scans. The model he trained on 488 patient images.
“Of the three models that performed best, all were based on handcrafted features,” she said. “None of them performed better than deep learning models.”
Dr. Elliot Siegel, professor and vice chair of research information systems at the University of Maryland, Baltimore, added that given the recent surge in AI publications, some would think Radiomics is over. However, he noted that Radiomics is very active and will continue to be active.
Compared to traditional radiomics techniques, deep learning models require very large datasets and are expensive and time consuming to develop, Siegel said. He cited a national lung screening trial developed as a training dataset for AI, enrolling 53,454 patients to compare CT and chest radiography approaches to detect cancer.
“It cost $250 million to assemble,” he said.
Another challenge in implementing deep learning in nuclear medicine is that images acquired using PET tend to lack texture to visualize tumors. The voxel volume of a tumor is determined by the image matrix size, and in nuclear medicine these sizes are relatively small compared to ultrasound, for example, he said. Siegel said Radiomics will benefit from these factors.
Ultimately, Siegel predicted that the radiomics era will continue to be strong and will greatly complement deep learning in the future.
“I think radiomics is going to be a fairly long-lasting approach for nuclear medicine,” he said.
Copyright © 2023 AuntMinnie.com
