CoLe-CNN+: Context learning - Convolutional neural network for COVID-19-Ground-Glass-Opacities detection and segmentation
Highlights
- ●
Single and completely automated workflow for accurate lung segmentation, COVID lesions detection and segmentation.
- ●
CNN producing four different segmentation mask that are combined in order to obtain a more accurate result.
- ●
Specially developed post-processing tool to improve specificity, reaching 99.6%.
- ●
Our method outperformed the state of the art reaching an average of 98.9% accuracy in COVID lesions segmentation.
Abstract
Background and objective
The most common tool for population-wide COVID-19 identification is the Reverse Transcription-Polymerase Chain Reaction test that detects the presence of the virus in the throat (or sputum) in swab samples. This test has a sensitivity between 59% and 71%. However, this test does not provide precise information regarding the extension of the pulmonary infection. Moreover, it has been proven that through the reading of a computed tomography (CT) scan, a clinician can provide a more complete perspective of the severity of the disease. Therefore, we propose a comprehensive system for fully-automated COVID-19 detection and lesion segmentation from CT scans, powered by deep learning strategies to support decision-making process for the diagnosis of COVID-19.
Methods
In the workflow proposed, the input CT image initially goes through lung delineation, then COVID-19 detection and finally lesion segmentation. The chosen neural network has a U-shaped architecture using a newly introduced Multiple Convolutional Layers structure, that produces a lung segmentation mask within a novel pipeline for direct COVID-19 detection and segmentation. In addition, we propose a customized loss function that guarantees an optimal balance on average between sensitivity and precision.
Results
Lungs’ segmentation results show a sensitivity near 99% and Dice-score of 97%. No false positives were observed in the detection network after 10 different runs with an average accuracy of 97.1%. The average accuracy for lesion segmentation was approximately 99%. Using UNet as a benchmark, we compared our results with several other techniques proposed in the literature, obtaining the largest improvement over the UNet outcomes.
Conclusions
The method proposed in this paper outperformed the state-of-the-art methods for COVID-19 lesion segmentation from CT images, and improved by 38.2% the results for F1-score of UNet. The high accuracy observed in this work opens up a wide range of possible applications of our algorithm in other fields related to medical image segmentation.