Helping Emergency Care Physicians Diagnose Sepsis and Bacterial Infections with Machine Learning and Vision

Sepsis is a life-threatening disease where the host's response to an infection leads to inflammation that may result in multiple organ failure and eventually death of the patient. Sepsis is associated with very high mortality rates that make early detection crucial for treatment. Microcirculation is the flow of blood in the smallest elements of the cardiovascular system, the capillary network, where the exchange of oxygen takes place and is pivotal in the maintenance of homeostasis. Improvements in microcirculation function after early resuscitation have been associated with reduction in subsequent organ failure.

In previous studies, microcirculation dysfunction was experimentally induced in animal models with drugs and other agents followed by Sidestream Dark Field (SDF) imaging. SDF imaging is a non-invasive imaging modality and has been used to track changes in the microcirculation on mucosal surfaces. SDF uses green polarized light with wavelength of 550 nm which is absorbed by hemoglobin and makes red blood cells visible. The distinguishing parameters and measurements associated to microcirculation include MFI (Microcirculation Flow Index), PVD (Perfused Vessel Density), TVD (Total Vessel Density) and PPV (Portion of Perfused Vessels). These measurements require identification of vessels in microcirculation video frames. Several software systems have been developed to analyze the microcirculation images and videos for these measurements, but often fail to provide desired results. Marking of hand engineered features on each frame of the video is required at some stage of analyses and is time consuming and not accurate.

To our knowledge, this is the first study that successfully classifies human microcirculation image data using a deep neural network. In this work, we implement a CNN to analyze microcirculatory dysfunction captured by dark field imaging in human patients and distinguish between septic and nonseptic images with high accuracies. 

We investigated outputs from the last convolutional layer in our architecture, and using a t-SNE embedding show that the representation learned by the classifier successfully differentiates the frames. Additionally, an unsupervised learning approach, independent of clinical labels, was used to investigate the feature space of the microcirculation image validation dataset and showed clustering images from non-septic and septic patients. These were first demonstrations in this field of separation between  learnable features in the data. 

We reason that the salient feature space used by our trained classifier has diagnostic applications for evaluating microcirculation dysfunction in humans. We also hypothesize that our neural network may be learning highly discernible features between images from septic and non-septic patients. The overlap, if any, between diagnostic features used by human experts and our machine learning algorithms is an active area of investigation in our research group. We also continue to investigate the learned feature space and it's explainability to human physicians.