Under-representation for Female Pelvis Cancers in Commercial Auto-segmentation Solutions and Open-source Imaging...

Area of Science:

• Medical imaging and radiation therapy
• Artificial intelligence in healthcare
• Gender representation in medical technology

Background:

• Artificial intelligence (AI) based auto-segmentation is increasingly used in radiation therapy (RT) workflows.
• Commercial solutions for organs at risk (OARs) and open-source datasets are vital for training AI algorithms.
• The equitable representation of female and male anatomies in these AI solutions is not well-studied.

Purpose of the Study:

• To investigate the representation of female and male anatomies in commercially available AI auto-segmentation solutions for OARs.
• To assess the availability of open-source imaging datasets for training AI algorithms across genders.
• To identify potential gender disparities in AI-driven RT tools.

Main Methods:

• Inquiries were sent to eight vendors about their gender-specific OAR auto-segmentation solutions.
• The Cancer Imaging Archive (TCIA) was screened for publicly available female and male anatomy imaging datasets.
• Data on solution availability, release dates, and OAR segmentation presence were analyzed.

Main Results:

• All vendors offered male pelvis AI auto-segmentation; 5/8 offered female pelvis solutions.
• Female breast and pelvis solutions were released later than male pelvis solutions (median 0.6 and 2.3 years, respectively).
• TCIA datasets predominantly featured male pelvis data with OAR segmentations, while female datasets lacked OAR segmentations.

Conclusions:

• Commercial AI auto-segmentation solutions and open-source datasets show a notable bias towards male anatomy.
• This gender disparity in AI tools and data may perpetuate inequities in radiation therapy.
• Addressing this imbalance is crucial for equitable cancer care across genders.