A Study to Validate and Improve an Automated Image Analysis Algorithm to Detect Tuberculosis in Sputum Smear Slides

Diascopic, Llc, Cleveland OH

Investigators

Linked publications, trials & patents

Abstract

Abstract/Summary In this SBIR, we propose to validate our handcrafted image analysis algorithm for auto-detecting Mycobacterium tuberculosis (MTB) in a digitized sputum smear. Once validated in a blinded study against manual microscopy and culture (the gold standard), we will try to improve our handcrafted algorithm by integrating, where appropriate, deep-learning approaches (via Convolutional Neural Networks (CNN)). Our novel diagnostic device (the Diascopic iON platform) uses automated image analysis to detect pathogens of interest. Through a blinded study (400 slides), we will assess the iON's effectiveness in detecting MTB. Our aim is to achieve >99% accuracy vs. microscopy, and sensitivity-specificity vs. culture of 80% and 99%, respectively. Currently, the iON platform can detect MTB on a Ziehl-Neelsen (ZN) stained sputum smear in less than 60 seconds, with accuracy of 95% vs. microscopy. The primary objective of this SBIR is to meet or exceed the minimal requirements for the WHO Target Product Profile (published 2014) of a rapid sputum-based test for detecting TB at the microscopy-center level of the health-care system. We will accomplish this feasibility study through a collaborative effort with the Case Western Reserve University-Uganda (UCRC) research team. A full-slide digitization and automated image analysis of 400 ZN slides is planned while on the ground in Uganda. Results will be published in an appropriate peer-reviewed journal for dissemination to the relevant TB pathology and provider community. A secondary objective of this SBIR is to improve our handcrafted algorithm through the use of deep- learning techniques (CNN). We will collaborate with Dr. Madabhushi (Case Western Reserve) - a world leader in Deep Learning methodologies ? on this portion of the study. We are optimistic that by combining our handcrafted approach with a deep-learning approach, we can identify MTB bacilli more effectively (i.e. faster and more accurately). We will leverage the lessons-learned in this study to develop algorithms for other developing-world diseases like Onchocerca (river blindness), Plasmodium (malaria), and Shistomes (schistosomaisis). Successful completion of this SBIR will show that the iON can truly become a platform for automated pathogen detection, which will shift lab practices toward faster & more standardized routines that are performed by unskilled workers. If we're successful in this Phase I SBIR, we will develop auto-detect algorithms for 3-4 other pathogens in a phase II SBIR. We will then market the iON platform to resource-limited clinics in countries adversely affected by developing-world diseases. It is our experience that such clinics are seeking a rapid, low cost, accurate and simple diagnostic tool to improve their efficiency and their ability to detect and treat diseases.