Under-Representation for Female Cancers in Commercial Auto-Segmentation Solutions and Open-Source Imaging Datasets

Abstract

Auto-segmentations methods to aid radiation therapy (RT) workflows have recently emerged with the increasing availability of commercial solutions for organs at risk (OARs) in addition to open-source imaging datasets that support training for new auto-segmentation algorithms. Here, we explored whether female and male cancer sites are equally represented among these solutions. Inquiries were sent out to five major RT vendors regarding their currently available auto-segmentation solutions. Additionally, The Cancer Imaging Archive (TCIA) was screened for publicly available imaging datasets pertaining to female and male tumor sites. The five commercial solutions provided a median of 103 (range: 60-120) OAR auto-segmentations of which the majority concerned the head and neck (45 (24-55)) and thorax (34 (27-43)) and were provided by all vendors (Table). Prostate as a site was also provided by all vendors and included 17 (9-20) auto-segmentations. A total of 23 publicly available TCIA imaging datasets involved the female anatomy (breast: 19; cervix: 2; ovarian: 1; uterus: 1) while 11 imaging datasets involved the male anatomy (prostate). No OARs segmentations were available for the 23 female-specific datasets while 27% of the 11 prostate datasets included segmented OARs. Three vendors and two TCIA datasets provided organs involved in the male sexual function apparatus (neurovascular bundle and penile bulb), whereas nipple or areola segmentations were not available among the commercial solutions for breast or among the TCIA breast datasets. None of the TCIA datasets or any of the five commercial solutions provided OARs for the female pelvis such as organs involved in reproduction (ovaries), sexual health (clitoris, vagina) or the cervix and uterus. Further, auto-segmentations provided for OARs trained exclusively on the male pelvis are likely inadequate for female cancers given the substantial anatomical differences between genders. Commercial auto-segmentation solutions and open-source imaging datasets together include considerably more datasets, tumor sites and consequently more OAR auto-segmentations pertaining to male cancers compared to female cancers. Despite a 1.4 times higher incidence for female cancers (breast: 300,590; female pelvis: 114,810; male cancer: 299,540; Siegel RL et al CA Cancer J Clin 2023), auto-segmentation models are lacking, and this gender disparity is likely to lead to suboptimal care for female-specific cancers.