P28 Weakly supervised deep learning on CT scans predicts survival from chronic pulmonary aspergillosis. (December 2018)

• Record Type: Journal Article
• Authors: Zusag, M; Angheleanu, R; Norhan, H; Nwankwo, L; Periselneris, J; Armstrong-James, D; Loebinger, MR; Desai, SR; Semple, T; Angelini, E; Shah, A
• Is Part Of: Thorax. Volume 73(2018)Supplement 4
• Page Start: A110
• Page End: A111
• Publication Date: 2018-12
• Subjects: Chest -- Diseases -- Periodicals; Thorax; Chest -- Diseases; Periodicals; Periodicals; 617.54
• DOI: 10.1136/thorax-2018-212555.186
• Languages: English
• ISSNs: 0040-6376

Abstract

Abstract : Background: Chronic Pulmonary Aspergillosis (CPA) is a severe fungal infection caused by the ubiquitous genus Aspergillus. Individuals with mild immunocompromise and/or pre-existing lung conditions are susceptible to CPA. Signs on high-resolution computed tomography (HRCT) imaging include uni- or bilateral fungal balls in lung cavities and associated pleural thickening and fibrosis. Despite antifungal therapy, mortality is high. Early diagnosis and radiological identification of disease progression are key to improve prognosis. We have designed a weakly-supervised deep learning network to recognize CPA, localize affected lung regions and predict 2 year survival post baseline on HRCT imaging. Methods: Our dataset consists of 75 normal HRCT studies and 277 studies from 99 patients showing signs of CPA, which were gathered over a period of 12 years. Following segmentation of the lung regions, an original approach was used via axial-view projection of the average intensity values (in HU) within the segmented lung masks, which provide contextual visualization of the whole lung volume in just two dimensions (figure 1b). We then fine-tuned the deep Convolutional Neural Network (CNN) VGG19 to classify axial-projections into two classes: CPA or not CPA. Data augmentation was performed via rotation of HRCT scans in 3D prior to projection, as well as scaling and rotating projections in 2D. The CNN was designed to output maps of visual features, which serve as a localizer of pathological signs on the studies (figure 1c). Finally, the CNN was trained to predict 2 year survival from baseline imaging. Results: For the classification of CPA, we achieved an accuracy on the training set (n=2790 projections) of 99.2\% and on the test set (n=962 projections) of 96.2\%. For 2 year survival prediction, we achieved an accuracy of 86.1\% on the training set (n=510 projections) and 85.8\% on the test set (n=255 projections). Conclusion: We present an original framework to simplify HRCT imaging via axial projections and exploit these in a deep-learning framework to enable accurate CPA identification and prediction of 2 year survival. Further work will aim to facilitate early CPA diagnosis and stratification and pave the way towards an automated scoring system with accurate mortality prediction.