Human-AI Interaction With AI-Assisted Tumor Overlays in Pediatric Whole-Body Magnetic Resonance Imaging: Exploratory...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Pediatric Radiology
• Oncology Surveillance

Background:

• Artificial intelligence (AI) offers potential for personalized clinical care, especially in radiology.
• Integrating AI into pediatric oncology workflows is challenging, despite the critical need for early cancer detection.
• Whole-body MRI (wbMRI) surveillance in Li-Fraumeni syndrome (LFS) presents an opportunity for AI-assisted tumor detection.

Purpose of the Study:

• To assess the feasibility of an AI overlay for highlighting potential tumors in pediatric surveillance wbMRI.
• To explore the impact of AI assistance on radiologist workflow, lesion marking, follow-up recommendations, and perceived workload.

Main Methods:

• A patch-based AI segmentation model was developed and trained on 675 pediatric LFS surveillance wbMRI scans.
• A reader study involved 2 radiologists independently reviewing wbMRI cases with and without AI assistance.
• Evaluation metrics included interpretation time, lesion marking, follow-up recommendations, and subjective feedback.

Main Results:

• AI assistance altered radiologist interpretation workflows, with mixed effects on efficiency and lesion detection.
• Average interpretation time per case increased with AI use for both radiologists.
• Subjective feedback indicated reduced stress with AI, but inter-reader variability highlighted the need for AI calibration.

Conclusions:

• AI-assisted wbMRI interpretation may enhance tumor detection in pediatric cancer surveillance, potentially reducing false negatives.
• Careful implementation is crucial to address workflow efficiency and inter-radiologist variability.
• Future clinical translation requires larger studies, AI model refinement, and user-centered design to build trust and improve outcomes.