A Novel Unsupervised Approach for COVID-19 Lung Lesion Detection Based on Object Completion Technique

Abstract

Automated segmentation of COVID-19 lesions from CT images is a prerequisite for quantitative assessment of the infections, enabling accurate and timely screening of the disease diagnosis/progression. Most of the deep learning-based approaches require a large dataset with known label maps for the lesions, which would be labor-intensive and highly time-consuming. In this paper, an innovative unsupervised framework is developed for COVID-19 infection segmentation from CT images that relies on a technique of object completion to detect lung lesions. To this end, two deep learning networks were separately trained in a supervised manner for patch-wise completion of lung volumes on 450 COVID-19 and 450 normal subjects (DL-Covid and DL-Norm, respectively). The artificial void areas were created within the lung volume to be predicted by the developed models. The DL-Covid and DL-Norm models were applied to the CT images of COVID-19 subjects. Since the DL-Covid was trained by the COVID-19 dataset, this model would have higher sensitivity to complete the missing regions with infection. In contrast, the DL-Norm model bears very low sensitivity to the infected areas (owing to the training with only normal samples). The final lesion probability map is obtained by subtracting the two predicted lung volumes. An external validation dataset, including 50 COVID-19 CT images with manual lesion segmentation (reference), was utilized for evaluating the performance of the proposed framework through the calculation of quantitative parameters such as Dice coefficients and Jaccard index (JC). The proposed method achieved an average Dice of 0.70 ± 0.034 and JC of 0.63 ± 0.062 for the COVID-19 infection segmentation. The promising results of the proposed method revealed its accuracy in predicting COVID-19 infections without the need for an annotated dataset. This approach could easily generate large noisy or weakly labeled datasets to train the dedicated deep learning models.

Ähnliche Arbeiten