Evaluation of an AI-Based TB AFB Smear Screening System for Laboratory Diagnosis on Routine Practice

Abstract

The most robust and economical method for laboratory diagnosis of tuberculosis (TB) is to identify mycobacteria acid-fast bacilli (AFB) under acid-fast staining, despite its disadvantages of low sensitivity and labor intensity. In recent years, artificial intelligence (AI) has been used in TB-smear microscopy to assist medical technologists with routine AFB smear microscopy. In this study, we evaluated the performance of a TB automated system consisting of a microscopic scanner and recognition program powered by artificial intelligence and machine learning. This AI-based system can detect AFB and classify the level from 0 to 4+. A total of 5930 smears were evaluated on the performance of this automatic system in identifying AFB in daily lab practice. At the first stage, 120 images were analyzed per smear, and the accuracy, sensitivity, and specificity were 91.3%, 60.0%, and 95.7%, respectively. In the second stage, 200 images were analyzed per smear, and the accuracy, sensitivity, and specificity were increased to 93.7%, 77.4%, and 96.6%. After removing disqualifying smears caused by poor staining quality and smear preparation, the accuracy, sensitivity, and specificity were improved to 95.2%, 85.7%, and 96.9%, respectively. Furthermore, the automated system recovered 85 positive smears initially identified as negative by manual screening. Our results suggested that the automated TB system could achieve higher sensitivity and laboratory efficiency than manual microscopy under the quality control of smear preparation. Automated TB smear screening systems can serve as a screening tool at the first screen before manual microcopy.

Retrospective validation of MetaSystems’ deep-learning-based digital microscopy platform with assistance compared to manual fluorescence microscopy for detection of mycobacteria

“…Comparable concordance rates of 92.7% ( 16 ) and 95.7% ( 26 ) with MM have been described in recent studies evaluating other AI-powered automated AFB microscopy systems in high TB incidence settings. Interestingly, Tomasello et al observed a drop in sensitivity (from 97.0% to 70.7%) when a similar version of the MetaSystems’ AFB detection software was employed with assistance at a DNN PT of 50% ( 27 ).…”

“…Meanwhile, with its current performance, detected AFB-smear-positive slides would require review by manual microscopy for confirmation and semi-quantification. The adoption of such technology within an established AFB testing algorithm could help streamline the use of molecular detection assays, where nucleic acid amplification tests would be reserved for AFB-smear-positive samples identified digitally ( 9 , 16 , 17 ). Overall, a-DM has the potential to improve laboratory productivity by allowing redistribution of the workforce particularly in high-throughput laboratories with low incidence TB.…”

Comparative analysis of five etiological detecting techniques for the positive rates in the diagnosis of tuberculous granuloma

“…The current study employed the N-acetyl-l-cysteine–NaOH (NALC–NaOH) technique to process specimens for both digestion and decontamination purposes [7] . Following centrifugation at 3200 × g for 20 min, the specimen was concentrated, and the precipitate was reconstituted with 0.067 Msterile phosphate buffer (pH 6.8).…”

A Novel Digitized Microscopic Images of ZN-Stained Sputum Smear and Its Classification Based on IUATLD Grades

“…However, the outcomes were categorized as AFB + or AFB-. Another assessment of an AI-based system for AFB detection was used in the same year 2022 by Hsiao et al in [27]. Each smear contained 200 images, which differed according to the WHO standard of at least 20-300 FoVs depending on tuberculosis level based on the IUATLD grades [31].…”

“…The CAD system can help pathologists to perform TB diagnostic smear microscopy more accurately and quickly. Two recent methods for sputum smear microscopy diagnosis have been developed to create CAD for TB: artificial intelligence (AI) [15], [22], [23], [24], [25] and mechanical automation [15], [26], [27], [28].These methods are similar to digital pathology, which uses an image recognition algorithm to detect AFB by digitally scanning the TB smear slides. These two combinations make the examination procedure more sensitive and accurate.…”