2026-02-09 スタンフォード大学
Prakash lab
<関連情報>
- https://news.stanford.edu/stories/2026/02/octopi-low-cost-autonomous-microscope-ai-malaria-diagnosis
- https://www.medrxiv.org/content/10.1101/2025.03.21.25324364v2
Octopi 2.0: AIを活用した診断アプリケーションのためのオープンでスケーラブルな顕微鏡プラットフォーム Octopi 2.0: Open and Scalable Microscopy Platform for AI-powered Diagnostic Applications
Hongquan Li, Heguang Lin, Pranav Shrestha, Rinni Bhansali, You Yan, Jaspreet Pannu, Kevin Marx, Wei Ouyang, Lucas Fuentes Valenzuela, Ethan Li, Anesta Kothari, Jerome Nowak, Hazel Soto-Montoya, Adil Jussupov, Maxime Voisin, Kajal Maran, Oswald Byaruhanga, Joaniter Nankabirwa, Bryan Greenhouse, Prasanna Jagannathan, Manu Prakash
medRXiv Posted: March 27, 2025
DOI:https://doi.org/10.1101/2025.03.21.25324364
Abstract
Access to quantitative, robust, and affordable diagnostic tools is essential to address the global burden of infectious diseases. While manual microscopy remains a cornerstone of diagnostic workflows due to its broad adaptability, it is labor-intensive and prone to human error. Recent advances in artificial intelligence (AI) and robotics offer opportunities to automate and enhance microscopy, enabling high-throughput, multi-disease diagnostics with minimal reliance on complex supply chains. However, current automated microscopy platforms are often costly and inflexible — barriers that are especially limiting in low-resource settings. Here we present Octopi 2.0, an open, highly configurable, general-purpose automated microscopy platform for a broad range of diagnostic applications, including sickle cell anemia and antibiotic resistance that we have reported recently. Applying Octopi to imaging malaria parasites with 4’,6-diamidino-2-phenylindole (DAPI) staining, we discovered a spectral shift in fluorescence emission that allows rapid screening of blood smears at low magnification with throughput on the order of 1 million blood cells per minute. We further developed image processing and deep learning-based segmentation and classification pipelines to enable real-time processing for malaria diagnosis. For real-world performance validation, we collected a data set of 213 clinical samples from Uganda and the United States with a total of 905 million red blood cells and around 1.4 million malaria parasites. Using a ResNet-18 model and only one round of retraining, the model is able to achieve on average less than 5 false positive parasites/µL and a per-parasite level false negative rate of less than 8% in our test dataset. This per-cell performance implies a limit of detection (LoD) around 12 parasites/µL, and we measured patient-level performance of >97% specificity and sensitivity in our independent test data set of clinical samples from 73 patients/donors. As more data is collected in larger validation studies, we expect the robustness and performance of the model to continue to improve according to what we observe in our proof-of-concept experiments carried out in this study. With significant cost reduction in hardware compared to current automated microscopes and an open and versatile approach for tackling multiple diseases with standard glass slide-based sample preparation, we envision Octopi 2.0 to help enable the “app store” for equitable data-driven, AI-powered diagnostics of many diseases and conditions.
