Integrated Smartphone-Based Computer Vision and Machine Learning Platform for Identification of Surgical Site Infections

Abstract

The infection of surgical wounds, also known as surgical site infections (SSI), represents a significant financial cost for health care systems worldwide but also represents a threat to the life and health of women in developing countries who give birth by Cesarean section (C-section). In order to help monitor and detect SSI in women who recently underwent C-section births, this thesis presents the design and development of an integrated smartphone application that can be used by community health workers (CHW) to help detect SSI using a smartphone camera image. This mobile application includes four main components: (1) a computer vision image capture algorithm with automated image scaling, cropping and rotation; (2) automated image quality assessment to provide real-time feedback to the CHW; (3) an image processing pipeline to perform image sampling, color correction, and brightness adjustment; (4) integrated image-based machine learning prediction, making use of a previouslydeveloped convolutional neural network (CNN) model. The integrated smartphone application, created with the Android Java SDK, is primarily designed to operate in rural parts of the world where there is a lack of Internet access. However, the mobile application is also designed to connect and synchronize data with a remote electronic medical record (EMR) server developed at MIT, known as the PyMed EMR server. In this thesis, I describe the design and implementation of the main components of the mobile application and the complete application work flow. I also discuss the performance of the application on different mobile phone models as well as the performance trade-off between online and offline wound infection prediction.