Segmentation of lung lobes and lesions in chest CT for the … – Nature.com

Datasets

Due to its retrospective nature, informed consent was waived, and all data were anonymized. This project was approved by the human research ethics committee of the Chulabhorn Research Institute (research project code 167/2564) and complied with the Declaration of Helsinki. These COVID-19 patients were confirmed by RT-PCR acquired from Chulabhorn Hospital who underwent non-contrast enhanced axial chest CT as a part of routine clinical care throughout the pandemic.

In this study, we randomly selected 124 cases from the database. The selection contained 28 cases without lung lesions and 96 cases with lung lesions. According to TSS, experienced radiologists classified the cases with lung lesions as mild, moderate, and severe. We divided the selection into 3 groups, i.e., training set, test set 1, and test set 2. The training set was used in model training and validation for lung segmentation and lesion segmentation; test set 1 was for segmentation performance evaluation; and test set 2 was for TSS prediction evaluation. We also randomly selected these cases for each group. In addition, for the training set and test set 1, the numbers of cases across different severity types were set to be equal to prevent class imbalance in the training set (the class imbalance causing a potential bias in the trained model) and for a fair comparison in test set 1. The number of CT slices in these cases ranged from 92 to 208. This information was described in Table 1.

The lung CT data were saved in JPEG format with a resolution of 512512 pixels and labeled by a program called LabelME41 (version 4.5.12). The resulting labels were in JavaScript Object Notation (JSON) format. All labeled data were validated by four radiologists and then converted into matrices for model training and evaluation.

In the data preprocessing phase, CT scan images (JPEG format) and labeled data (JSON format) were resized using the cv2.resize function from 512512 pixels to 256256 pixels to minimize the required memory resources (RAM). The interpolation parameter was set to INTER_AREA for the CT scan images and INTER_NEAREST" for the label data because this solution prevented any alteration of the values specified in each pixel. In addition, our model input shape was fixed at a size of 128256256. The CT volumes were adjusted to 128 images per patient according to the following three conditions (Fig.1).

The first condition, if the CT volume comprised 128 or fewer images, a 256256 zero-padding matrix was added to increase the volume to 128 images.

The second condition, if the CT volume had between 129 and 175 CT images, 128 images from the CT volume's middle range were selected to train the model because both the lung parenchyma and lesions appear in this range.

The third condition, if the CT volume contained more than 175 CT images, we skipped the CT slice by selecting only odd-numbered images and adding a 256256 zero-padding matrix to reach a total of 128 images.

Overall pipeline of the data preprocessing.

A color adjustment method was applied to improve image contrast by using the contrast-limited adaptive histogram equalization (CLAHE) technique42, which is available in the OpenCV library43. The CLAHE parameters were set to a clipLimit of 3 and a tileGridSize of (8, 8). Models were trained/tested in two experiments: the first with original images (no color adjustment) and the second with CLAHE-adjusted images.

A 256-slice dual-energy CT scanner (Revolution CT with Gemstone Spectral Imaging (GSI) Xtream, GE Healthcare) at Chulabhorn Hospital was used in this study. An axial chest CT scan without contrast agent was applied. The protocol started with a scout view from lung apices to lung bases in anteriorposterior (AP) and lateral views, and followed by an axial chest scan covering lung apices through bases from inferior to superior. The parameters were quiet breath inspiration, 1.25mm thickness, 0.28s/rotation, 0.992 pitch, GSI calculated kVp, 190mA, lung window of (1550, 700), soft tissue window of (400, 40), and postprocessing multiplanar reconstruction. The scan time was less than 1.6s.

Two models were used in this study: (1) a lung lobe segmentation model and (2) a lesion segmentation model. Training set: 32 cases were split into 24 cases (75%) for model training and 8 cases (25%) for validation, where the dataset was divided equally at each severity type to prevent overfitting. According to related studies, a model that combines a 3D-UNet structure with a DenseNet or ResNet is effective in segmenting parts of the image precisely. Therefore, pre-trained DenseNet121, DenseNet169, DenseNet201, ResNet18, ResNet34, ResNet50, ResNet101, and ResNet152 models were obtained through a segmentation-models-3D package from Solovyev et al.44.

Lung lobe segmentation. A multiclass semantic segmentation model was used to segment the five lung lobes. Annotated labels consisted of six categories: 0, 1, 2, 3, 4, and 5, which indicate the background, right upper lobe (RUL), right lower lobe (RLL), right middle lobe (RML), left upper lobe (LUL), and left lower lobe (LLL), respectively.

Lesion segmentation. The lesion model was developed from a binary semantic segmentation model that outputs the value 1 for lesion areas and 0 for background areas. Images without extrapulmonary regions are preferred for lesion model training. The dataset used for model training was preprocessed as described in the data preparation section.

In the model training process, lung lobe and lesions segmentation models were trained on servers equipped with an Intel(R) Xeon(R) Gold 6126 CPU at 2.60GHz, 40GB of RAM, and an NVIDIA Tesla V100 SXM2 GPU. Figure2 shows the overall workflow. The models output is the predicted classfor each pixel, which is then used to compute the percentage area of lesions in each lung lobe for the CT score. This score is then used to calculate the TSS value for diagnosing the severity of the current pathology. For both models, Adam optimization was used, the loss function was a hybrid loss function (focal loss+Dice loss), the learning rate was set to 0.0001, a regularizer that applies L2 regularization was used with a value of 0.01, the batch size was set to 1, and the maximum number of epochs was 200. The lesion model activation function was set to sigmoid with a dropout rate of 0.4, whereas the pulmonary lobe model activation function was set to SoftMax with a dropout rate of 0.2. The hybrid loss technique45, which combined focal loss and Dice loss, was used to improve model performance.

Segmentation Model Workflow and Total Severity Score Calculation Protocol for Lung CT Scans.

The PI in each lung lobe was calculated by dividing the number of predicted lesion pixels by the total number of lung lobe pixels in the CT volume. The predicted lesion pixels were obtained from the output of the lesion segmentation model, whereas the predicted lung lobe pixels were derived from the output of the lung lobe segmentation model, in which the value of each pixel identifies the lobar type in the lung CT image. Therefore, the PI was calculated by performing the following equation.

$$Percentage, of{, Infection}_{lobe}= frac{Lesion, Area, (pixels)}{Lung ,Lobe, Area, (pixels)}times 100$$

The TSS proposed by Chung et al.20 was calculated from the sum of the five-lobe CT score, which was calculated from PI based on the criteria listed in Table 2. The severity of COVID-19 patients can be classified from the TSS value based on the severity criteria in Table 3 and the following equation:

$$TSS =CT ,Scor{e}_{RUL}+CT ,Scor{e}_{RML}+CT ,Scor{e}_{RLL}+CT, Scor{e}_{LUL}+CT, Scor{e}_{LLL}$$

The most commonly utilized measurement to evaluate image segmentation is the Dice similarity coefficient (DSC). The DSC calculated the relative overlap between the predicted area and ground truth, and it was used to choose the most appropriate model. The DSC was defined as follows:

$$DSC = frac{2TP}{2TP + FP + FN}$$

where the term true positive (TP) refers to an outcome such that the model correctly predicts the positive class, false positive (FP) is an outcome such that the model incorrectly predicts the positive class, and false negative (FN) is an outcome such that the model incorrectly predicts the negative class.

Hausdorff distance was proposed by Felix Hausdorff in 191446,47. The measure was applied to evaluate the models performance by measuring the distance between two images in pixels. The distance is defined as

$$Hleft(A,Bright)=mathrm{max}left(hleft(A,Bright), hleft(B,Aright)right)$$

$$hleft(A,Bright)={mathrm{max}}_{mathit{aepsilon A}}{mathrm{min}}_{bin B}||a-b||,$$

where A is a set containing p points (pixels on image A): ({{a}_{1},{a}_{2},dots ,{a}_{p}}) and B is a set containing q points (pixels on image B): ({{b}_{1},{b}_{2},dots ,{b}_{q}}). For implementation, we applied the function implemented in SciPy package48. Since the images used in this research were 256256 pixels, the Hausdorff distance range was ([0, 256sqrt{2}]).

Go here to see the original:

Segmentation of lung lobes and lesions in chest CT for the ... - Nature.com