AI-based mobile application to fight antibiotic resistance

Abstract

Antimicrobial resistance is a major global health threat and its development is promoted by antibiotic misuse. While disk diffusion antibiotic susceptibility testing (AST, also called antibiogram) is broadly used to test for antibiotic resistance in bacterial infections, it faces strong criticism because of inter-operator variability and the complexity of interpretative reading. Automatic reading systems address these issues, but are not always adapted or available to resource-limited settings. We present an artificial intelligence (AI)-based, offline smartphone application for antibiogram analysis. The application captures images with the phone's camera, and the user is guided throughout the analysis on the same device by a user-friendly graphical interface. An embedded expert system validates the coherence of the antibiogram data and provides interpreted results. The fully automatic measurement procedure of our application's reading system achieves an overall agreement of 90% on susceptibility categorization against a hospital-standard automatic system and 98% against manual measurement (gold standard), with reduced inter-operator variability. The application's performance showed that the automatic reading of antibiotic resistance testing is entirely feasible on a smartphone. Moreover our application is suited for resource-limited settings, and therefore has the potential to significantly increase patients' access to AST worldwide.

Fig. 1. Analysis of an AST plate with the App.

A prepared and incubated Petri dish (a) is positioned in a simple image acquisition setup made of cardboard (b), we used two containers available in the laboratory as stands. A picture of the plate is taken with a smartphone and the analysis follows the workflow described in (c): the Petri dish image is cropped and the antibiotic disks are found (c1); the image of each antibiotic disk is fed to a ML model that identifies the antibiotic (c2); the diameter of the inhibition zone is measured (c3) with an original algorithm. Finally, the Expert System uses the diameters to output interpreted results (c4).

Fig. 2. Screenshots of the App in action.

a The App displays a zoomed image of an inhibition zone and indicates with a dashed circle the automatically measured diameter and the detected antibiotic. The user can edit the results with the controls below the image. b The application can ask the users if they see the peculiar shapes of inhibition zones associated with certain resistance mechanisms. c At the end of the analysis, the interpreted results are shown to the user.

Fig. 3. Problematic images and the role of intensity contrast.

A few problematic images have been identified in the datasets. These correspond to damaged plates (a) and images with very poor visible contrast between the bacteria and the inhibition (b). Some inhibition zones are hard to isolate, even by eye. For comparison, a standard image looks like (c). The coupled effect of bacteria pigmentation and variable illumination produces a considerable variability in the bacteria-to-inhibition intensity contrast (a,b,c). The histogram in (d) shows the distribution of image contrast for standard and problematic images in AST set A1 (the contrast is defined here as the difference between the central intensity level of bacteria and inhibition): problematic images (in red) are a small fraction of the total, mainly concentrated in the lower contrast region. Finally, e shows the observed mean diameter difference in millimeters versus contrast (Data are presented as mean values ± SD): low contrast images yield worse results.

Fig. 4. Benchmark results on dataset A3.

The histograms (a,b) show the distribution of the absolute diameter differences between the App’s automatic procedure (auto) and the manual measurement with ruler (a) as well as with the diameter adjusted on the smartphone by the technicians (assisted, b). On the right, the heat-maps show the average absolute measurement difference among the eight technicians (given two readers i and j, square i, j represents the average difference between them) measuring with the ruler (c) and in assisted mode with the App (d). The assisted measure seems to reduce inter-operator variability.

Fig. 5. Distribution of species in the AST groups used in this study.

a Dataset A1. b Dataset A2. c Dataset A3.

Fig. 6. Variability in antibiogram pictures.

Examples of difficult cases for diameter reading (a). Non-circular inhibition shape (a1). Total or no inhibition (a1). Light reflections (a3). Colonies within the zone of inhibition (a4). Double inhibition zones (a5). Hazy borders (a6). Inhibition zone overlap and plate borders (a7). Low contrast (a8) defined as the difference between the inhibition and bacteria intensity value, compared to a high contrast (a9) image observed in dataset A1. The histograms in (b) show the contrast variability observed in the benchmark datasets (contrast is defined as the difference between the central gray levels of bacteria and inhibition, and normalized to the maximum available gray level). Observed variability in dominant hue (c).

Fig. 7. Characteristic shapes of inhibition zones due to resistance mechanisms.

a D-shape inhibition zone due to induction. b Appearance of inhibition between antibiotic disks due to synergy.