Use of Convolutional Neural Nets and Transfer Learning for Prediction of Surgical Site Infection from Color Images

One of the greatest concerns in post-operative
care is the infection of the surgical wound. Such infections are
a particular concern in global health and low-resource areas,
where microbial antibiotic resistance is often common. In
order to help address this problem, there is a great interest in
developing simple tools for early detection of surgical wounds.
Motivated by this need, we describe the development of two
Convolutional Neural Net (CNN) models designed to detect an
infection in a surgical wound using a color image taken from a
mobile device. These models were developed using image data
collected from a clinical study with 572 women in Rural
Rwanda, who underwent Cesarean section surgery and had
photos taken approximately 10 days after surgery. Infected
wounds (N=62) were diagnosed by a trained doctor through a
physical exam. In our model development, we observed a
trade-off between AUC accuracy and sensitivity, and we chose
to optimize for sensitivity, to match its use as a screening tool.
Our naïve CNN model, with a limited number of convolutions
and parameters, achieved median AUC = 0.655, true positive
rate sensitivity = 0.75, specificity = 0.58, classification accuracy
= 0.86. The second CNN model, developed with transfer
learning using the Resnet50 architecture, produced a median
AUC = 0.639 sensitivity = 0.92, specificity = 0.18, and
classification accuracy 0.82. We discuss the specific training
and optimization methods used to compensate for significant
class imbalance and maximize sensitivity.