Deep learning-based image processing methods to guide treatment of bacterial infections and combat antibiotic resistance
– Erik Hallström

Treatment of bacterial infections usually has to start before the responsible species and its antibiotic susceptibility are known. Conventional diagnostics require time-consuming bacterial culturing, and patients with severe infections are therefore given broad-spectrum antibiotics, a treatment that contributes to the development of antibiotic resistance.

This thesis investigates whether deep-learning-based image analysis can shorten the time to a diagnosis. The work follows two tracks.
In the first track, bacteria are loaded into the narrow traps of a microfluidic device and imaged with phase-contrast time-lapse microscopy as they grow and divide. Deep learning models are trained to recognize the species from these images.

Paper I shows that four species can be recognized, and that video models using several frames outperform single-image models, particularly at low spatial resolution.

Paper III extends the task to seven clinically relevant species and examines how performance depends on temporal and spatial resolution, training set size, augmentation and data quality, and on the texture, morphology and division patterns visible in the images.

Paper IV addresses the detection of bacteria in processed blood, where debris can resemble bacterial cells, and shows that temporal information reduces false positives. Paper V combines detection, phenotypic antibiotic susceptibility testing and species identification into a single workflow starting from clinical isolates not seen during training. Paper VI introduces a self-supervised pretraining task for time-lapse imaging data and asks whether autofluorescence adds useful information.

The second track concerns automatic analysis of bacterial cultures on agar plates. Paper II presents CombiANT Reader that analyses photographs of CombiANT, a novel antibiotic combination test. It segments bacterial growth zones and quantifies antibiotic interactions. It was evaluated on photographs of 100 plates taken by three users with different experience levels, and agreed closely with manual grading.

The results show that deep learning models can extract spatiotemporal patterns from phase-contrast images that are difficult for humans to interpret. If further developed and validated, faster bacterial detection, identification and susceptibility testing could allow clinicians to select antibiotics earlier, reduce unnecessary broad-spectrum treatment, help combat antibiotic resistance and in the long run, save lives.

About Erik Hallström