P–105 Clinical validation of mojo AISA, an artificial intelligence robotic CASA system. (6th August 2021)

• Record Type: Journal Article
• Authors: Monteiro, M; Thomas, D; Maillot, R; Simon, Z; Björndahl, L; Flanagan, J; Taha, M
• Is Part Of: Human reproduction. Volume 36:Supplement 1(2021)
• Publication Date: 2021-08-06
• Subjects: Human reproduction -- Periodicals; 618
• DOI: 10.1093/humrep/deab130.104
• Languages: English
• ISSNs: 0268-1161

Abstract

Abstract: Study question: Can a CASA system based on Artificial Intelligence perform as well as manual semen assessment, within the WHO error margins? Summary answer: The AI-based CASA systems that mimic high quality assessments show great potential for reducing clinical workloads while increasing treatment efficacy. What is known already: The field of male-factor fertility investigation is still lacking an automated semen analysis system that can be widely clinically adopted. By leveraging state-of-the-art robotics and Artificial Intelligence (AI), it was possible to build mojo AISA which is an AI and robotic platform designed according to WHO recommendation for semen analysis. This system is based on AI software with a unique convolutional neural network (CNN) that detects and measures sperm concentration and motility while ruling out unwanted cells and debris in raw samples. Study design, size, duration: This study presents and validates the mojo AISA device. A total of 60 patient samples at ANOVA Karolinska University Hospital were collected and results from manual assessment were compared to mojo AISA for concentration and motility. Semen samples were assessed manually (WHO 2010) and concurrently with Mojo AISA. Manual measurements ranged from 1–206M/ml. This study lasted from May 2020 to December 2020 following informed consent and ethics committee practices of ANOVA. Participants/materials, setting, methods: Sample preparation protocol for mojo AISA consisted of placing two 10ul drops and covering with two 22x22mm coverslip. Manual assessment followed ANOVA EQA procedures akin to the WHO. A CNN was trained using videos captured with mojo AISA as input data. Images were annotated to form a validation set by which the AI was trained. To account for sampling error, videos of Hamilton Thorne Accubeads+ were captured using mojo AISA and the mojo counting chambers. Main results and the role of chance: Comparing the concentration measured by mojo AISA with the known value for each microbead, results are in agreement of 86%, within the confidence interval of the microbeads. The mean relative error was 6.7% and maximum error was 11%. Therefore, Accubeads+ validation proved no observational error regarding the use of mojo AISA microscope. As for comparing mojo AISA to manual assessment for concentration, Pearson (Spearman) correlation was 0.95 (0.97). The mean relative error was 24.8% and maximum relative error was 71.1%, where 90% of samples were below 50% error. By looking at the concentration range between 10 and 20 M/ml, mojo AISA displayed a mean error of 18.5%. For motility, as comparing mojo AISA to manual assessment, a result of 35.4% mean relative error was obtained. To conclude, mojo's robotic solution shows promise for clinical practice as the AI continues to improve. In 6 months, sperm concentration correlation improved by 3-fold. Next, the AI will be further clinically trained for low concentration. Limitations, reasons for caution: mojo AISA requires further development, especially for very low concentration ranges, below 5M/ml, due to high sensibility to false positive detections. The same applies to post-vasectomy samples. Additionally, the necessity to compute the motility of each sperm scales poorly with high concentration generating a poor experience for high volume clinics. Wider implications of the findings: Automation is crucial in several industries. It enables fertility clinics & andrologists to standardize male factor infertility measurements (if paired with widespread standardization of protocols for automation) while enabling them to put more focus on demanding activities of their profession and removes human biases of inter-laboratory performance. Trial registration number: Not applicable