Convolutional Neural Net Models and Image Processing Methods for Predicting Surgical Site Infection

Abstract

Surgical site infections are an important cause of disability, cost, and even mortality, especially in low-resource settings, where patients have limited access to clinical facilities or trained medical professionals. As an alternative to a hospital-based diagnosis by a doctor, there is interest in making use of community health workers to help identify infections in patients’ homes. As an aid for diagnosis, we propose using mobile phone devices to capture an image of a wound and then apply a convolutional neural network (CNN) model to identify features of infection. For this thesis, I have explored both RGB images captured using a mobile phone camera and also thermal images captured using an external thermal camera module. The data for this work was collected as part of clinical studies conducted in rural Rwanda by Harvard University, consisting of two datasets: Dataset A (60 infected, 500 non-infected), and Dataset B (70 infected, 1,100 non-infected). From these datasets, separate na¨ıve CNN and transfer learning CNN models were constructed. The overall median AUC values for each model, based on the ROC curve, were as follows: Na¨ıve CNN for Dataset A (Median AUC = 0.65), Transfer learning CNN for Data A (Median AUC = 0.64), Na¨ıve CNN for Dataset B (Median AUC = 0.68), Transfer learning CNN for Data B (Median AUC = 0.86), Na¨ıve CNN for Thermal imaging (Median AUC = 0.86), Transfer learning CNN for thermal imaging (Median AUC = 0.90). In addition to model development, an image pre-processing pipeline was also developed through an extensive series of experiments to study the effect of image blur, pixel resolution, and color calibration. The performance of our models compares favorably to prior work done in the field of wound infection prediction, and to our knowledge, this is the first reported work using thermal imaging to predict infection. These results demonstrate that prediction of surgical infection is feasible using mobile phone imaging tools; it is hoped that this work can lead to new methods for identification of surgical site infection in low-resource areas as well as for outpatient care in developed countries.