02/14/2022
By Terence Griffin

The Kennedy College of Sciences, Department of Computer Science, invites you to attend a doctoral dissertation defense by Terence Griffin on "Deep Learning Applied to Tuberculosis Screening."

Candidate Name: Terence Griffin
Defense Date: Wednesday, March 2, 2022
Time: 1 to 2 p.m.
Location: This will be a virtual defense via Zoom
Dissertation Title: Deep Learning Applied to Tuberculosis Screening

Committee Chairs (Advisers):

  • Yu Cao, Professor, Computer Science Department, University of Massachusetts Lowell
  • Benyuan Liu, Professor, Computer Science Department, University of Massachusetts Lowell

Committee Member:

  • Yan Luo, Professor, Department of Electrical and Computer Engineering, University of Massachusetts Lowell


Brief Abstract:

Tuberculosis (TB) is a contagious disease affecting millions of people annually worldwide. TB primarily affects the lungs and is spread through the air when people cough, sneeze, or spit. Providing healthcare professionals with better information, at a faster pace, is essential for combating this disease, especially in Low and Middle Income Countries with resource-constrained health systems. In this work we investigate the use of deep learning approaches to the problem of TB screening.

Two large TB datasets are used for training our models. Combining the two datasets required the development of a preprocessing stage which includes lung segmentation and contrast enhancement along with a set of data augmentations. We show that these steps are necessary to combine the datasets and do not negatively affect a model's performance.

We describe how convolution neural networks (CNNs) trained using an object level annotated dataset of chest X-rays (CXRs) allows us to identify the location of pulmonary issues indicative of TB. We compare the performance of Faster R-CNN, Mask R-CNN, Cascade versions of each, and SOLOv2, demonstrating reasonable results with a well annotated dataset. We present a method to reduce the false positive rate by comparing the location of a detected object with the known location of areas where the detected class is likely to occur in the lung. Our results show that object detection and instance segmentation of CXRs can be achieved with a dataset of high-quality, object level annotations, and could be used as part of an automated TB screening process.

We then present eRxNet, a pipeline of CNNs designed to provide healthcare professionals with detailed and accurate analysis of CXRs for TB screening. The pipeline combines whole image classification, object detection (bounding boxes), and instance segmentation (polygonal masks) to provide data analysis at varying levels of detail. In order to construct a high performing system, a comparison of different CNN architectures applied to these tasks is presented. Images from two large TB datasets, UML-Peru and TBX11K, were used for training and evaluation of the models. We show that the resulting four stage pipeline of CNNs, using a combination of DenseNet, Faster R-CNN, and Mask R-CNN, has sufficiently strong performance to be a useful tool for TB screening.

To gain confidence in the ability for eRxNet to generalize beyond the datasets used for training we evaluated the component models using additional CXR datasets. Each of these tests show only a slight degradation in performance compared with the original datasets.

The eRx Repository is an online application to aid in the training of healthcare professionals in diagnosing TB. This tool provides access to the UML-Peru dataset, including expert annotations and selected results from eRxNet. It is hoped that this tool can help to address the shortage of trained professionals combating TB.