Artificial Intelligence Assisted Triage of Cervical Biopsies Using the Low-Cost, 3D-printed OpenFlexure Microscope: a Paired Comparative Feasibility Study

Objective Artificial intelligence has been utilized in pathology to analyze whole slide images for triage of biopsies for pathologist review. However, this has been done with prohibitively costly slide scanners that may be unavailable in low- and middle- income countries where the number of trained pathologists is limited. We evaluated whether a 3D-printed slide scanner (the OpenFlexure Microscope) could create images of sufficient quality that could be used for artificial intelligence triage of cervical biopsies and compared them to images created by a commercial slide scanner. Methods Cervical biopsy slides from 275 specimens at the University of Michigan were scanned by the OpenFlexure Microscope and Aperio GT450, where 186 were used for the model training/validation set, and the remaining 89 for the hold out set. Post capture edits including pixel resolution standardization and color correction were performed to ensure minimal differences between the two images. Our team created and tested five models using open-source software to classify biopsies into groups that were at high-risk and low-risk for high grade squamous intraepithelial neoplasia. Results The hold-out set contained 29 high-risk and 60 low-risk cases. Metrics ranged across models: area under the receiver operating characteristic curve 0.71-0.80, accuracy 0.64-0.72, precision 0.45-0.55, F1 score 0.45-0.64, and sensitivity 0.45-0.76. The 95% confidence interval for most of the metrics overlapped across all five models, indicating lack of statistical power to support any one model outperforming the others. Conclusion This comparative feasibility study suggests that an open-source, 3D-printed slide scanning device can produce images of sufficient quality for artificial intelligence-assisted cervical biopsy triage to aid pathologists in flagging high-risk cases needing expediting review. Additional research is needed with larger multi-site cohorts from low- and middle- income countries where tissue processing and disease prevalence differ to assess real world application.