RESUMEN
Coronavirus Disease 2019 (COVID-19) spreads around the world, seriously affecting people's health. Computed tomography (CT) images contain rich semantic information as an auxiliary diagnosis method. However, the automatic segmentation of COVID-19 lesions in CT images faces several challenges, including inconsistency in size and shape of the lesion, the high variability of the lesion, and the low contrast of pixel values between the lesion and normal tissue surrounding the lesion. Therefore, this paper proposes a Fully Feature Fusion Based Neural Network for COVID-19 Lesion Segmentation in CT Images (F3-Net). F3-Net uses an encoder-decoder architecture. In F3-Net, the Multiple Scale Module (MSM) can sense features of different scales, and Dense Path Module (DPM) is used to eliminate the semantic gap between features. The Attention Fusion Module (AFM) is the attention module, which can better fuse the multiple features. Furthermore, we proposed an improved loss function L o s s C o v i d - B C E that pays more attention to the lesions based on the prior knowledge of the distribution of COVID-19 lesions in the lungs. Finally, we verified the superior performance of F3-Net on a COVID-19 segmentation dataset, experiments demonstrate that the proposed model can segment COVID-19 lesions more accurately in CT images than benchmarks of state of the art.
RESUMEN
BACKGROUND: The COVID-19 disease is putting unprecedented pressure on the global healthcare system. The CT (computed tomography) examination as a auxiliary confirmed diagnostic method can help clinicians quickly detect lesions locations of COVID-19 once screening by PCR test. Furthermore, the lesion subtypes classification plays a critical role in the consequent treatment decision. Identifying the subtypes of lesions accurately can help doctors discover changes in lesions in time and better assess the severity of COVID-19. METHOD: The most four typical lesion subtypes of COVID-19 are discussed in this paper, which are GGO (ground-glass opacity), cord, solid and subsolid. A computer-aided diagnosis approach of lesion subtype is proposed in this paper. The radiomics data of lesions are segmented from COVID-19 patients CT images with diagnosis and lesions annotations by radiologists. Then the three-dimensional texture descriptors are applied on the volume data of lesions as well as shape and first-order features. The massive feature data are selected by HAFS (hybrid adaptive feature selection) algorithm and a classification model is trained at the same time. The classifier is used to predict lesion subtypes as side decision information for radiologists. RESULTS: There are 3734 lesions extracted from the dataset with 319 patients collection and then 189 radiomics features are obtained finally. The random forest classifier is trained with data augmentation that the number of different subtypes of lesions is imbalanced in initial dataset. The experimental results show that the accuracy of the four subtypes of lesions is (93.06%, 96.84%, 99.58%, and 94.30%), the recall is (95.52%, 91.58%, 95.80% and 80.75%) and the f-score is (93.84%, 92.37%, 95.47%, and 84.42%). CONCLUSION: The three-dimensional radiomics features used in this paper can better express the high-level information of COVID-19 lesions in CT slices. HAFS method aggregates the results of multiple feature selection algorithms intersects with traditional methods to filter out redundant features more accurately. After selection, the subtype of COVID-19 lesion can be judged by inputting the features into the RF (random forest) model, which can help clinicians more accurately identify the subtypes of COVID-19 lesions and provide help for further research.