Abstracts

RETURN TO ABSTRACT LISTING


E1596. Automated Detection of COVID-19 From Chest CT Scans Using Deep Convolutional Neural Networks
Authors
  1. Amit Kharat; DeepTek Imaging Private Limited
  2. Rohit Lokwani; DeepTek Imaging Private Limited
  3. Priyam Choudhury; DeepTek Imaging Private Limited
  4. Ashrika Gaikwad; DeepTek Imaging Private Limited
  5. Viraj Kulkarni; DeepTek Imaging Private Limited
  6. Aniruddha Pant; DeepTek Imaging Private Limited
Objective:
The COVID-19 pandemic continues to spread and wreak havoc across the globe. The complex process of sample collection, analysis, delayed reporting and irregularity in test techniques of RT-PCR; having a high specificity but variable sensitivity for test results,(1) can be overcome by automated detection through chest Computed Tomography (CT). CT scans have diagnosed patients with early disease, recovered disease, without symptoms and even with negative RT-PCR test results (2). Therefore, purpose of our prospective study, was to detect patients for COVID-19 pneumonia, by applying machine learning algorithm to chest CT images, and natural language processing to radiologist reports.

Materials and Methods:
834 chest CTs were taken from the institution and public database, and 275 scans were annotated by VIA (VGG Image annotator), outlined and marked COVID-19 positive, by expert radiologists, based on typical findings like ground-glass opacities, consolidations, bilateral and peripheral, “crazy paving” (3) and “reverse halo” signs. The image resolution was kept at 512 x 512 pixels and split into training (657 slices), validation (120 slices) and test (266 slices) data sets. A 2D segmentation model was built using the slices from these scans. The architecture used was U-net with Xception-net as the encoder, which was optimized using binary cross-entropy as loss function. We used transfer learning, by fine tuning a model pre-trained on chest X-rays using weights from the model built for COVID detection in chest X-rays as transfer learning has been proven to have better efficacy than random initialization.

Results:
The model, was run on test data set of 230 CT scans and demonstrated a sensitivity of 0.963 (95% CI: 0.94-0.98) and specificity of 0.936 (95% CI: 0.92-0.95). Further, it was evaluated on 140 CT scans from three different sources like Italian, Chinese and Indian private hospitals, which demonstrated sensitivity of 0.964 (95% CI: 0.88-1) and specificity of 0.884 (95% CI: 0.82-0.94).

Conclusion:
Chest CT has proved to have higher sensitivity than RT-PCR (4) and can be used as screening during the pandemic. The convolutional neural network (CNN) trained in this data achieved promising results for detection of COVID-19 vs non-COVID-19. Currently, our 2D model is built at slice level, and given that a CT image has thousands of slices, it adds to time complexity of processing data. Therefore, in the future, we plan to implement a 3D model that will take the chest CT scan as input, and give out masks for the infected areas, and also differentiate between COVID-19 and various pneumonias.