From Sparse Radiologist Annotations to Precise Image-Guided Interventions

When it comes to implementing precise radiation therapy plans for tumors that move with respiration, such as those in the lungs or abdomen, 4D CT remains the tool of choice. 

“By capturing a tumor’s full range of motion, 4D CT can help radiologists define a more accurate target volume and minimize damage to surrounding healthy tissues,” explained Xiaoman Zhang, PhD, a post-doc student at Harvard University. “The problem is that 4D CT contouring is a particularly time-consuming, labor-intensive process.” 

During a Sunday session, Dr. Zhang presented the research of Yi Luo, who is currently pursuing a PhD in biomedical engineering at Johns Hopkins University School of Medicine in Baltimore.

“Not only do 4D CT scans acquire multiple 3D CT volumes, with each volume representing a different phase of the breathing cycle, radiologists must also analyze the tumor and surrounding organs across all phases before making an IGTV (internal gross tumor volumes) contour for efficiency consideration,” said Luo, who spoke to Daily Bulletin about his research prior to RSNA.  

According to Luo, although automatic segmentation methods are available, most require dense, expert-level annotation, meaning that 4D CT contouring is often done manually.  

Believing that there had to be a better way, Luo decided to look at whether existing weak annotations, such as isocenters (ISO), could guide automatic lung tumor segmentation on 4D CT. 

“ISO points are routinely created in clinical workflows but have rarely been used in 4D CT,” Luo remarked. “What we wanted to know was whether a single click ISO, together with a powerful foundation model, could replace the labor-intensive task of 4D CT contouring?”

Turning Free Treatment ISO into an Effective Foundation Model Prompt 

To find out, Luo helped develop a novel segmentation workflow where radiologist-placed ISO markers serve as weak annotations to initialize a MedSAM2 foundation AI-model for 4D CT image segmentation. 

Evaluated on a cohort of 42 lung cancer patients who had previously undergone radiotherapy at Johns Hopkins Hospital, the proposed segmentation workflow achieved a mean Dice coefficient of 0.5705±0.2264 across the test cohort. The average processing time was 114.04±41.22 seconds per patient, demonstrating notable computational efficiency. 

When analysis was restricted to conventionally used clinical phases (phases 0, 50, 60), the Dice coefficient was 0.5503±0.2702, while sensitivity notably increased to 0.6239±0.2173. “This study turns the free treatment ISO, which is already present in routine care, into an effective prompt for a foundation model, achieving 4D lung-tumor segmentation with no extra clinician effort,” Luo said.

To understand the impact each individual respiratory phase has on overall segmentation performance, the study also performed a correlation analysis, with the omittance of phases 0 or 90 from the input data resulting in the most significant decline in performance. 

“These findings underscore that while phase 0 is routinely included in 4D CT tumor analysis, phase 90 is also a potentially critical phase for comprehensive assessment,” Luo noted.

Reducing Workloads and Variability, Improving Treatment Efficacy and Safety

By creating a practical pathway for converting routine clinical markers into AI-ready weak supervision, this new approach to segmentation has the potential to reduce contouring workload, variability, and turnaround time in 4D CT interpretation and radiotherapy planning. 

“Faster, more consistent tumor delineation can accelerate treatment-planning cycles, meaning therapy can start earlier, while improved segmentation accuracy can reduce geographic misses and spare normal tissues, thus enhancing treatment efficacy and safety,” Luo concluded.

Access the presentation, “From Sparse Radiologist Annotations to Precise Image-Guided Interventions: An Efficient MEDSAM2-Enhanced Strategy for Lung Tumor Segmentation on 4D CT,” (S2-SSRO01-3) on demand at RSNA.org/MeetingCentral