Use of Crowd Innovation to Develop an Artificial Intelligence-Based Solution for Radiation Therapy Targeting

Area of Science:

• Medical Physics
• Artificial Intelligence
• Oncology

Background:

• Radiation therapy (RT) is crucial for cancer treatment, but a shortage of radiation oncologists limits global access.
• Accurate tumor segmentation is vital for effective RT planning but is time-consuming and prone to human error.

Purpose of the Study:

• To assess if crowd innovation can quickly generate artificial intelligence (AI) solutions for lung tumor segmentation in RT.
• To determine if these AI solutions can match the accuracy of expert radiation oncologists.

Main Methods:

• A 10-week, online, prize-based challenge involving 564 contestants was conducted.
• A curated dataset of 461 patients' CT scans with expert lung tumor segmentations was used.
• AI algorithms were developed and evaluated on an independent dataset, benchmarked against human expert variation.

Main Results:

• The top 5 AI algorithms, combined via an ensemble model, achieved an accuracy (Dice coefficient = 0.79) comparable to expert interobserver variation.
• Algorithm performance significantly improved across three phases, with the final ensemble model reaching an S score of 0.68.
• The challenge successfully generated AI tools that replicate expert physician skills in tumor segmentation.

Conclusions:

• Crowd innovation combined with AI rapidly produced automated algorithms for critical RT tasks.
• These AI solutions have the potential to improve global cancer care by disseminating expert clinical skills.
• AI can help overcome workforce limitations in radiation oncology, especially in resource-limited settings.