A new artificial intelligence tool developed by researchers at Northwestern Medicine has matched expert physicians in mapping lung tumors for radiation therapy and may even improve on their accuracy by identifying tumor regions some doctors miss. The study detailing this tool, called iSeg, was published NPJ Precision Oncology.
“We’re one step closer to cancer treatments that are even more precise than any of us imagined just a decade ago,” said senior author Mohamed Abazeed, MD, PhD, chair and professor of radiation oncology at Northwestern University Feinberg School of Medicine.
Tumor segmentation is integral to radiation therapy and involves defining the precise boundaries of a tumor on CT scans. It determines where and how radiation will be targeted to destroy cancer cells while sparing healthy tissue. Segmentation is currently performed manually, which can be time-consuming, prone to variation from person to person, and subject to oversight.
Unlike previous segmentation algorithms developed from static images, iSeg is the first 3D deep learning tool that can provide motion-resolved tumor segmentation to capture the way lung tumors shift with each breath, information that is important for planning the most precise and effective radiotherapy for each patient.
To develop iSeg, the Northwestern team tapped a large dataset of CT scans and physician-delineated tumor contours from 739 lung cancer patients from nine clinics within the Northwestern Medicine and Cleveland Clinic health systems.
The researchers trained a 3D UNet deep neural network to segment gross tumor volumes (GTVs) on 4D CT scans, enabling it to generate internal target volumes (ITVs) that take into account respiratory motion. Validation of the tool was conducted on two independent cohorts of 161 and 102 patients. iSeg’s segmentations closely matched expert-drawn contours, achieving a median Dice Similarity Coefficient (DSC) of 0.73 internally and comparable scores in the external datasets (DSC = 0.70 and 0.71). “iSeg approximated inter-physician variability, achieving DSC values near to the human-to-human benchmark (~0.80),” the researchers wrote.
In addition, the tool often identified additional tumor regions not segmented by physicians. These regions, when left untreated, were statistically linked to higher local failure rates. As the researchers noted, “higher false positive voxel rate (regions segmented by the machine but not the human) were associated with increased local failure (HR: 1.01 per voxel, p = 0.03), suggesting the clinical relevance of these discordant regions.”
This study builds on decades of research documenting variability in manual tumor delineation, particularly in regard to the use of complex treatment modalities such as stereotactic body radiotherapy (SBRT). “Manual delineation remains prone to consequential inter-observer variability, especially for ultra-conformal techniques like SBRT,” the researchers wrote. Earlier studies have shown such variability can affect dosimetry leading to worse patient outcomes.
In the development of iSeg, the researchers also sought to detail not only its segmentation abilities but also its clinical potential. “Unlike prior approaches, iSeg also integrates clinical outcome correlation and motion-resolved segmentation, two critical advancements for practical clinical translation,” the researchers wrote.
While iSeg has shown great potential in these early tests, the researcher acknowledged some areas where it didn’t perform as well. “Its segmentation performance, while approaching human benchmarks, remains modestly lower than expert consensus in certain cases, particularly for small, low-contrast central tumors adjacent to complex anatomy,” the researchers wrote. In addition, the saliency maps used to visualize AI decision-making are sensitive to changes in input and may lack boundary precision.
Despite these areas, iSeg is now undergoing prospective testing in clinical settings, where its output is being compared in real-time to physicians’ work. “This technology can help support more consistent care across institutions, and we believe clinical deployment could be possible within a couple of years,” said study co-author Troy Teo, PhD, instructor of radiation oncology at Feinberg.
Additional improvements to the iSeg include adapting the tool to other cancers that are treated with radiation therapy like liver, brain and prostate cancer and training it on additional imaging techniques such are MRI and PET scans. The team is also designing a study to compare physician-only, AI-only, and combined segmentation approaches.
