Label-Free Segmentation of COVID-19 Lesions in Lung CT

Abstract

Scarcity of annotated images hampers the building of automated solution for reliable COVID-19 diagnosis and evaluation from CT. To alleviate the burden of data annotation, we herein present a label-free approach for segmenting COVID-19 lesions in CT via voxel-level anomaly modeling that mines out the relevant knowledge from normal CT lung scans. Our modeling is inspired by the observation that the parts of tracheae and vessels, which lay in the high-intensity range where lesions belong to, exhibit strong patterns. To facilitate the learning of such patterns at a voxel level, we synthesize 'lesions' using a set of simple operations and insert the synthesized 'lesions' into normal CT lung scans to form training pairs, from which we learn a normalcy-recognizing network (NormNet) that recognizes normal tissues and separate them from possible COVID-19 lesions. Our experiments on three different public datasets validate the effectiveness of NormNet, which conspicuously outperforms a variety of unsupervised anomaly detection (UAD) methods.

Domain-Adaptive Artificial Intelligence-Based Model for Personalized Diagnosis of Trivial Lesions Related to COVID-19 in Chest Computed Tomography Scans

“…In conclusion, these comparative results (Table 4) highlight the superior performance of our model over all baseline models [22,29,[44][45][46]. Moreover, there are some existing studies [12][13][14]41,[47][48][49][50] that provide state-of-theart benchmarks for our selected datasets. Therefore, we also compared the results of our methods with those of these methods [12][13][14]41,[47][48][49][50], which are given in Table 5.…”

“…Moreover, in a t-test analysis (proposed vs. second-best methods), we obtained an average p-value less than 0.05 (to be specific, average p-value = 0.044) that implies the discriminative performance of our model against [41,47,50] with a 95% confidence score. In conclusion, these comparative results highlight the superiority of our method over all the existing methods [12][13][14]41,[47][48][49][50] related to the segmentation of COVID-19 lesions using chest CT scans.…”

“…Moreover, there are some existing studies [12][13][14]41,[47][48][49][50] that provide state-of-theart benchmarks for our selected datasets. Therefore, we also compared the results of our methods with those of these methods [12][13][14]41,[47][48][49][50], which are given in Table 5. First, Zhang et al [12] proposed a new variant of C-GAN, called CoSinGAN, with the capability to be learned from a single image and synthesize high-quality CT images for efficient training of a segmentation model.…”

“…In the case of the cross-dataset, Ma et al's [14] given benchmarks showed better results than [12]. Subsequently, in [41,47,49,50], different segmentation methods (i.e., Miniseg [41], GASNet [47], Label-Free [49], and DASC-Net [50]) were proposed to achieve state-of-the-art results for effective segmentation of COVID-19 lesions in chest CT scans. Based on the results given in Table 5, we find the existing GASNet [47], Miniseg [41], and DASC-Net [50] as the second-best methods in case of COVID-19-CT-Seg (Exp#1), MosMed (Exp#2) and the cross dataset (Exp#3 (fold 1)), respectively.…”

Recognition of COVID-19 from CT Scans Using Two-Stage Deep-Learning-Based Approach: CNR-IEMN

“…Since the deep learning approach has proved its efficiency in most computer vision tasks [11] (including medical imaging tasks [3,12]), most of the approaches that have been proposed for COVID-19 analysis from CT scans are CNN-based approaches. In general, the approaches that use CT scans for COVID-19 analysis can be classified into: segmentation [13][14][15] and recognition [16][17][18][19][20] approaches.…”