A new generation of diagnostic systems available at the point of care (POC) could save lives and reduce the spread of infectious diseases worldwide through early detection and treatment.
Optical microscopy remains the gold standard for the diagnosis of many diseases; however, its accuracy is dependent on the availability and expertise of the analyst at the POC. This limitation is increased by the dependence on labor-intensive examination processes, lack of standardization, high interobserver variability, insufficient precision in sample quantification and, as a consequence, a high misdiagnosis rate.
One promising solution is the use of artificial intelligence algorithms to aid diagnosis, though these systems often require large datasets, which are not always accessible. In our project we propose a data-efficient AI-based methodology for disease diagnosis that focuses on the sample rather than the targeted disease, enabling the identification of multiple diseases, including skin Neglected Tropical Diseases. By leveraging a methodology for classifying multiple species using self-supervised learning (SSL), model performance can be significantly improved with limited annotated data.
As a result, we developed a foundational model based on over 100K microscope images (10x, 40x and 100x magnification, from four study sites) acquired from 332 patients. From this database, more than 89K images were used for SSL pretraining, while the remaining images were labeled by experts and used for fine tuning with an 80%-20% patient-level split. Using ViT as a transformer based model with DINO as the SSL strategy, we achieved 95% accuracy (F1 Score) across the different species with just 100 labels per new class. This shows that further species can be included with a very limited number of labels.
Our work presents a generalized AI framework for detecting and classifying multiple diseases within a single model, that can be extended to other sample types and deployed into an autonomous microscope. This solution can be integrated into smartphones, facilitating real-time diagnosis and universal monitoring of infectious diseases.
Elena Dacal*, David Bermejo-Peláez*, Lin Lin*, Lucia Pastor, Roberto
Mancebo, Ramón Vallés-López, Daniel Cuadrado, Claudia Carmona,
Alexandra Martín Ramírez, Victor Anton Berenguer, María Flores-Chavez,
José M. Rubio, Miguel Luengo-Oroz
*co-first authors