Our research toward the development of these algorithms has resulted in novel image analysis algorithms (graph cut, atlas based) that identify lung boundaries and delineate lung regions in the posteroanterior CXR. The research has also resulted in novel combinations of shape, edge and texture descriptors computed from pixels within the lung boundaries. These image descriptors are then used to train supervised machine learning classifiers (e.g., Convolutional Neural Network-based (CNN) Deep Learning, SVM). We have acquired 4 datasets for testing these algorithms and have made them publicly available for enabling scientific research on the topic. These datasets have been downloaded by over 250 researchers from academic and industrial research labs worldwide. Performance of our lung segmentation algorithm has been shown to be over 95% accurate. In addition, the classification accuracy for TB detection has been shown to be 88% which is measured as area under the receiver operating characteristic (ROC) curve. The curve measures the response of the classifier at various operating points and indicates the trade-off between sensitivity and specificity of its performance. Both results have been published in high quality archival journals. This year we advanced our machine learning methods to include CNN-based Deep Learning models that have advanced classification performance significantly. We arebeta-testing our machine learning models in the field to differentiate between normal CXR and those exhibiting lung diseases. The implemented system, which received an HHS Innovates Award, is installed in a truck equipped with a mobile x-ray unit and is deployed by a Kenyan-NGO (Academic Model for Providing Access to Healthcare, AMPATH) and is traveling through numerous sites in rural western Kenya, visiting scores of patients weekly. Research continues with acquiring digital CXR from the field for further training and testing of the algorithms, and the implementation of deep learning techniques (CNN) for fine-tuned classification of the CXR.
| Karargyris, Alexandros; Siegelman, Jenifer; Tzortzis, Dimitris et al. (2016) Combination of texture and shape features to detect pulmonary abnormalities in digital chest X-rays. Int J Comput Assist Radiol Surg 11:99-106 |
