We retrospectively collected data from patients diagnosed with PTB who were admitted to four hospitals from January 2010 to December 2021. As sputum smear result is associated with the infectivity of PTB (18), we used sputum smear result to assess infectivity. According to the sputum tuberculosis smear interpretation criteria in the WS 288-2017 (19), the sputum smear result was categorized into six grades from low to high: negative (–), weakly positive (±), positive (+), positive (2+), positive (3+), and positive (4+). A negative result (–) indicates the absence of acid-fast bacilli in 50 consecutive microscopic fields. A weakly positive result (±) indicates the presence of 1–9 acid-fast bacilli in 50 microscopic fields. The positive (+) category refers to 10–49 acid-fast bacilli found in 50 microscopic fields. positive (2+) indicates 1–9 acid-fast bacilli found in each microscopic field, and positive (3+) indicates 10–99 acid-fast bacilli found in each microscopic field. The positive (4+) indicates more than 100 acid-fast bacilli found in each microscopic field. The report for positive (2+) should be based on the observation of a minimum of 50 fields, while for positive (3+) and higher positive results, a minimum of 20 fields should be observed.

We defined the negative group as having three or more negative sputum smear results within a month, and no positive results within the next 3 months, and this indicates weak infectivity. The positive group was defined as having at least one-time positive sputum smear results within a month, indicating relatively strong infectivity. We selected the highest number of multiple sputum smear results for each person in the positive group as the grading criterion. The National Health Commission of the People's Republic of China's diagnostic standards for PTB served as the foundation for the diagnosis of PTB (WS 288-2017) (19). After the patients were grouped according to the sputum smear results, chest CT images of the patients in DICOM format were collected and matched. The interval between sputum smear tests and CT image acquisition was <30 days, and patients with unclear PTB diagnoses or poor-quality lung CT images were excluded.

Figure 1 displays a detailed flowchart of the procedure for gathering data. We included one CT scan image from each patient, totalling 925 CT scan images. Of these, 591 images were split for training, with 118 images set aside as a validation set for model parameter selection. The remaining 334 images were used for testing. Hospital 1 provided the images for the training and validation sets. Hospitals 2, 3, and 4 provided the images for the external test set.

The patients were positioned prone and instructed to inhale and hold their breath as much as possible during the lung field scan. The visual field was adjusted to fit the size of each patient. All patients were scanned using spiral CT scanners following the same protocol. The CT images have an in-plane pixel spacing of 5 mm, an in-plane resolution of 512 × 512, and the number of slices ranges from 47 to 70. All CT data were converted from the original DICOM format to the NIFTI format to ensure data desensitization.

Figure 2 shows the overview of our PTB infectious distinguish method. The first step is lung segmentation. Considering that contagious PTB occurs in the lung parenchyma, lung region segmentation can make the model focus on the lung without interference from other areas, thus reducing the difficulty of the analysis. Lungmask (20) is an open-source lung segmentation model based on deep learning, which is used to perform automatic segmentation of lungs on 3D CT images.

Based on the segmented lung mask, a CT image is divided into left and right lung images. Then the left and right lung images are projected from three sides, that is, the mean and standard deviation of pixels in three directions are calculated. Finally, the images of all planes are scaled to a uniform size, as shown in Figure 3. The advantage of projecting 3D images to 2D images is that the 2D CNN can be used to analyze these data. Compared with 3D CNN, 2D CNN is lighter and easier to train. Besides, 2D projection increases the number of samples, because a 3D CT image is converted into three 2D projection samples in different directions, which can alleviate the over-fitting of our model.

In this study, we proposed a PTB infectious convolutional neural network (CNN) which can be called TBINet for rapid identification of contagious people with PTB. As shown in Figure 4, the TBINet contains two ResNet (21) backbones used to extract the features of left and right lung projection images, respectively. The details of the ResNet backbone are also shown in Figure 4, which consists of convolution (Conv) layers, batch normalization (BN) layers, rectified linear units (ReLU), residual blocks, max pool layer, and adaptive avgpool layer. Then the extracted left and right lung features are fused by maximum operation. Finally, the fused features pass through a full connection layer to get the final infectious prediction score. The prediction score is a value between 0 and 1. All the prediction scores can be divided into two groups by setting the cut-off value, with those above the cut-off value being positive and those below being negative.

A loss function is typically necessary for CNN optimization. For TBINet optimization, we employ a binary cross-entropy loss function, which is specified as follows:

Where N is the number of samples, y_i is the label of the sample i, and p_i is the score of the sample i predicted by our TBINet.

The following describes a few implementation specifics. Our model was trained for 300 epochs using the Adam (22) optimizer and a step-decay learning rate. The learning rate was 0.0001 at the beginning. Some straightforward online data augmentation techniques were applied to the training set to reduce overfitting, such as random flipping, rotating, and zooming.

The dataset was split into three sets: a training set, which was used to train the TBINet; a validation set, which was used for model parameter selection; and an external test set, which was used to evaluate the generalization ability of our model. We contrast our TBINet with the existing deep learning-based PTB classification methods, including 3D ResNet (14), AECNN (11), and LightCN (12), to demonstrate the advantages of our approach. The area under the receiver-operating-characteristic curve (AUC), accuracy, sensitivity, specificity, precision, and F1 score were calculated for these models to be evaluated and compared in the validation and external test set. The confusion matrices were also computed to display the prediction results of all compared methods.

The age difference was compared using the t-test, and gender was evaluated using the Chi-square test. The 95% confidence interval (CI) of the AUC metric was calculated for model evaluation.

SPSS statistical software version 26.0 (IBM Corp., Armonk, NY) and Python software version 3.6.6 (Python Software Foundation, Wilmington, DE, USA) were used for all analyses and model construction. All statistical tests were 2-sided, and P < 0.05 were considered to be statistically significant.

This study included images from 925 in-patients. Both the training and validation sets were obtained from Hospital 1 (Table 1), whereas the external testing dataset was collected from Hospitals 2–4. There were 721 (77.9%) males and 204 (22.1%) females with a mean age of 50.6 ± 17.3 years. The ratio of positive to negative was 1.92, and training, validation, and testing sets were set at 51.1, 12.8, and 36.1% of the full set, respectively. There were no significant differences in the sex ratio between positive and negative groups (P = 0.432) and age (P = 0.192; see Table 2).

Figure 5 shows the receiver-operating-characteristic curves (ROCs) of all compared methods. The 3D ResNet achieves the highest AUC (0.928) on the training set, but a lower AUC on the validation (0.796) and external test set (0.714), which indicates that it has serious over-fitting. Our TBINet achieves the best performance on the validation and external test sets with AUC of 0.817 and 0.754, respectively, which shows the superior generalization ability of our model. The detailed comparisons of AUC, accuracy, sensitivity, specificity, precision, and F1 score on the validation and external test sets are listed in Tables 3, 4, respectively. The results show that the proposed TBINet achieves the best performance with all metrics on the validation and external test set. As shown in Figure 6, the confusion matrices also show the predictions of the TBINet have fewer false positives and false negatives.

Gradient weighted class activation mapping (Grad-CAM) is a commonly used tool for CNN explainability. It uses gradients of a specific target that flow through the convolutional network to localize and highlight regions of the target in the image. Grad-CAM can reveal the areas of the image that the model relies on to make positive or negative predictions. Figure 7 presents some examples of Grad-CAM in action on our TBINet model.

By observing and analyzing the image areas activated by Grad-CAM, we discovered: in the negative group with weak infectivity, the lesions were characterized by fibrous proliferation foci, bronchiectasis, pleural thickening, stretching, and adhesion, with a few cases exhibiting cavities; in the positive group with strong infectivity, the lesions were primarily characterized by exudate, and some cases exhibited caseous pneumonia, accompanied by cavity formation. These findings are consistent with the pathological features of pulmonary tuberculosis (PTB) (22).