Acute PTB often presents with acute respiratory symptoms, including dry cough, fever, and chest pain (Moreira et al., 2011). Symptom duration generally surpasses 2 weeks before hospitalization. The clinical appearance is comparable to non-tuberculous community-acquired pneumonia (CAP), but patients with PTB experience less intense pleural unease, toxemia, and fatigue than those with non-tuberculous bacterial pneumonia (Sharma and Mohan, 2004). Acute PTB is more likely to cause weight loss than non-tuberculous CAP, while hemoptysis is relatively rare (Morena et al., 2023). The typical symptoms of TB are vividly shown in Figure 1. Generally speaking, the diagnosis of PTB is made using a combination of conventional and contemporary techniques. This diagnosis depends primarily on the patient’s medical symptoms and findings from the physical examination, complemented by various diagnostic test results. These include bacteriological tests, the tuberculin skin test (TST), imaging studies such as X-ray or CT scans, histopathological evaluations, T-SPOT test, and the reaction to the therapy plan of antituberculosis medications (Kang et al., 2021). The distribution of mycobacterium TB in the human body is shown in Figure 4.

The WHO regards TB culture as the “gold standard” for diagnosing TB; nonetheless, conventional solid and liquid culture media exhibit several drawbacks, such as impracticality, bacterial cross-contamination, and extended culture durations. Sputum smear microscopy is a diagnostic technique for TB that is widely acknowledged as effective (Kessel et al., 2023, pp. 2013–2017). Due to its affordability and relative simplicity compared to other advanced diagnostic methods, sputum smear microscopy remains a critical diagnostic instrument for PTB, particularly in nations with low incomes. In this process, sputum samples are coughed up by patients exhibiting symptoms and are treated with chemicals and applied to plain glass microscope slides. Yet the inconsistent clinical performance of this method, combined with the challenges in sputum collection from patients and accessibility to healthcare services, constitutes one of the main reasons for TB being undiagnosed. Subsequently, these slides are subjected to laboratory analysis to detect the presence of TB. The resulting images from a sputum smear test are typically viewed using fluorescence or light microscopy. The resolution and size of these images are determined by the level of magnification employed. For laryngeal swabs, the pooled sensitivity was 57.8% (95% CI: 50.5–65.0), and the specificity was 93.8% (88.4–96.8). For nasopharyngeal aspirates, the sensitivity was 65.2% (95% CI: 52.0–76.4), and the specificity was 97.9% (95% CI: 96.0–99.0). For oral swabs, the sensitivity was 56.7% (95% CI: 44.3–68.2), and the specificity was 91.3% (95% CI: 81.0–96.3) (Savage et al., 2023). Fluorescence microscopy offers advantages such as labor reduction and enhanced productivity. Nevertheless, the potential drawback of this technique is the danger of false-positive results, which is a result of the fluorochrome dyes’ non-specific binding (Steingart et al., 2006). The non-invasive collection process and its association with TB transmission have long been why breath has been regarded as an appealing diagnostic specimen for TB (Ghosh et al., 2021; Pham and Beauchamp, 2021). Electronic nose tests reportedly have an estimated sensitivity of 92% (95% CI: 82–97%) (Saktiawati et al., 2019).

Serological assays frequently demonstrate inadequate sensitivity and specificity, relying on a humoral immune reaction to detect antibodies against TB antigens (Steingart et al., 2011). A strategy attracting increased attention is the identification of host responses indicative of TB infection. In this regard, interferon-gamma release assays (IGRAs), such as T-Spot (Oxford Immunotec) and QuantiFERON (Qiagen), have limited use in detecting acute infections but are effective in detecting latent TB. However, IGRAs are affected by diseases such as diabetes, and the high cost also limits its application in underdeveloped regions (Takasaki et al., 2018). Currently, the T-SPOT test is extensively utilized for diagnosing infections caused by TB (Sollai et al., 2014). Despite its importance in identifying Mtb infections, a significant drawback of the T-SPOT test is that it cannot differentiate between active TB and latent TB infection (Ling et al., 2013).

The benefits and constraints of the QuantiFERON-TB Gold (QFT) test parallel those of the T-SPOT test. The QFT test offers a more straightforward operational process than T-SPOT, eliminating the need to separate peripheral blood mononuclear cells and opting to use whole blood instead (Lalvani and Pareek, 2010). Only a handful of antigenic biomarkers have been identified for TB. Lipoglycan lipoarabinomannan (LAM) is the most extensively investigated, most promising, and accessible from a simple sample such as urine (MacLean et al., 2019). The FujiLAM is a lateral flow urine test that finds LAM antigens. In adults with TB, it has a sensitivity of 70% and a specificity of 93% (Li Z. et al., 2021). The TST is a traditional diagnostic technique that uses a pure protein derivative of tuberculin to identify delayed-type hypersensitivity responses. Individuals infected with TB can generate sensitized T cells that recognize MTB antigens. Upon re-stimulation by MTB antigens, these sensitized T cells secrete different soluble lymphokines that enhance permeability, local erythema, and induration (Kowalewicz-Kulbat et al., 2018). The TST method finds the average diameter of the induration 72 h after an injection of a pure protein derivative tuberculin. An induration diameter <5 mm or no reaction is negative; ≥5 mm is positive (Abubakar et al., 2018).

Molecular imaging, which integrates molecular biology with medical imaging, is increasingly being explored to enhance our understanding of PTB. This high-tech imaging method allows biological processes at the molecular and cellular levels inside live things to be witnessed and measured. Techniques such as single-photon emission computed tomography (SPECT) are being investigated for their potential to offer detailed molecular-level data. SPECT’s high sensitivity and resolution make it a promising tool for identifying and tracking disease processes associated with PTB (Dimastromatteo et al., 2018). Along with smear microscopy, the Centers for Disease Control and Prevention suggests that each individual who might have PTB should also have at least one nucleic acid amplification (NAA) test, like polymerase chain reaction (PCR) (Parrish and Carroll, 2011). The high guanine and cytosine in the TB genome make it harder to conduct PCR studies. As a result, when dealing with MTB, it is essential to meticulously consider the methods for sample collection, bacterial cell disruption, nucleic acid isolation, and PCR assay design. The “Xpert MTB/RIF assay” is a rapid NAA test capable of detecting TB and determining resistance to rifampicin (Naidoo et al., 2023). The sensitivity of Xpert MTB/RIF Ultra when applied to oral swabs varied between 45% and 77.8%, in contrast to an approximate sensitivity of 90% for sputum samples (Mesman et al., 2019; Lima et al., 2020; Andama et al., 2022). Identifying bacterial RNA allows for the pinpointing of active TB. The quantification of bacterial quantity in sputum samples can be achieved using the detection of 16S ribosomal RNA, providing a sensitivity comparable to that of solid culture techniques (Honeyborne et al., 2014). Loop-mediated isothermal amplification (LAMP) is a PCR method that functions at a uniform temperature. The TB-LAMP method exhibits a sensitivity marginally lower than the Xpert MTB/RIF assay, yet both tests maintain similar specificity (Phetsuksiri et al., 2020). The WHO advocates TB-LAMP as a superior alternative to smear microscopy because of its improved diagnostic efficacy (Huang et al., 2022). A line probe test (LPA) is a quick, accurate, and flexible way to determine if someone has TB. It can detect TB in various clinical specimens, including sputum, pleural, and cerebrospinal fluid. The LPA can identify resistance to first-line TB medications such as isoniazid. Different commercial LPA kits are available, including the GenoType MTBDRplus 1.0 assay from Hain Lifescience and the INNO-LiPA Rif TB kit offered by Innogenetics (Rossau et al., 1997; Crudu et al., 2012). LPA is crucial for the management of multidrug-resistant TB. Truenat MTB, Truenat MTB Plus, and Truenat MTB-Rif Dx assays are quick molecular real-time PCR tests that can find TB. Results are usually ready in an hour (Nikam et al., 2013; Georghiou et al., 2021).

Most individuals with PTB exhibit abnormal chest X-ray (CXR) findings, which indicate a PTB diagnosis (Nachiappan et al., 2017). Given the relative accessibility of CXR and its utility in identifying these signs, the WHO advises using chest radiography for TB screening in populations at high risk for the disease (Liang et al., 2022). Different TB lesions appear differently on X-rays; exudative lesions appear as cloud-like or patchy shadows, proliferative lesions as nodular shadows, and caseous lesions as high-density and uneven shadows. The advantages of CT scanning include its relatively low cost, enhanced capability to differentiate between tissue types, rapid image acquisition, and broader accessibility. Compared to X-ray chest films, CT scans provide more explicit sectional images of the lungs, avoiding the issue of overlapping pictures and displaying the delicate structures and lesion details of lung tissue. Various pathological changes of TB, such as exudation, proliferation, caseation, fibrosis, and calcification, can be well displayed on CT scans. CT can detect early small lesions, bronchial dissemination foci, and mediastinal lymph node enlargement, aiding in a definitive diagnosis. It is also useful in diagnosing suspected TB cases that are atypical or negative on X-ray chest films. Additionally, CT has an extra role in assisting with fluid aspiration, biopsy confirmation, and guiding therapeutic interventional procedures, such as fluid drainage. PET/CT, which utilizes the cellular uptake of ¹⁸F-fluorodeoxyglucose to assess pulmonary inflammation, is highly sensitive for the early detection of TB (Ankrah et al., 2018). X-ray chests are the initial step in investigating PTB, followed by ultrasound (US), CT, and magnetic MRI for further evaluation. Additional imaging modalities, such as intravenous urography and barium studies, may also be helpful.

Innovative techniques are being developed to evaluate, track, and quantify PTB conditions at the point of care. Techniques such as lung ultrasonography and electrical impedance tomography are becoming more popular as they complement traditional diagnostic methods. These methods are being extensively researched for their potential to supplement standard procedures and, in certain respiratory conditions, to serve as an alternative due to the absence of ionizing radiation and their simplicity (Ball et al., 2017). A meta-analysis has revealed that chest US, when used for diagnosing pediatrics PTB, has a sensitivity of 84% and a specificity of 38% (Muljadi et al., 2024). Many molecular biomarkers can also be used to diagnose PTB. The combined sensitivity and specificity of monocyte to lymphocyte ratio (MLR) in detecting TB are 79.5% (95% CI: 68.5–87.3) and 80.2% (95% CI: 67.3–88.9), respectively (Adane et al., 2022).

Non-sputum biomarker tests for TB could have a market value of between $56 million and $84 million in countries with a high incidence of TB, like South Africa, Brazil, China, and India. It is thought that 14 million tests will be done (Tb Diagnostics Market Analysis Consortium, 2014, 2015; Maheshwari et al., 2016; Zhao et al., 2016). Dai et al. (2019) made a model for predicting TB by testing three iron-related biomarkers in blood serum, which can recognize TB. There is optimism that this approach could be broadened to enhance the diagnostic techniques for PTB. Next-generation sequencing is considered a revolutionary method for medication susceptibility testing of TB, providing data much faster than conventional clinical culture-based methods (Walker et al., 2015). Mass spectrometry technology can accurately detect biomarkers, which helps with the early diagnosis of TB. Chen et al. (2020) utilized Matrix-Assisted Laser Desorption/Ionization Time-of-Flight Mass Spectrometry and Liquid Chromatography-Tandem Mass Spectrometry to identify a TB-specific serum peptide signature, thereby creating diagnostic models for swift and accurate TB detection. More details can be found in Table 1, which compares various TB diagnostic methods.

Deep learning neural networks, a category of computational models, can learn complex feature hierarchies by deriving advanced features from simpler ones. Fukushima introduced this concept in 1980, inspired by the mechanisms of human vision based on biological principles (Yoon and Kim, 2020). DL emulates the human brain’s process of information filtering to facilitate accurate decision-making. DL instructs a computational model to handle inputs through a series of layers, analogous to the human brain’s approach, to bolster data prediction and categorization. Each layer feeds its output to the subsequent layer, akin to the progressive filtering mechanisms employed by neural networks within the brain. The iterative feedback process continues until the production remains consistent with the previous iteration. Weights are first allocated to each layer to produce the desired output, and these weights are further refined during the training process to attain the exact result (Lalmuanawma et al., 2020).

In the fields of CT-based disease diagnosis, multiple DL models are frequently utilized for tasks. CNNs like VGGNet, Google Net, ResNet, and DenseNet are essential for image categorization and feature extraction (Pavithra et al., 2023). U-Net, V-Net, SegNet, and DeepLab are commonly employed for segmentation jobs because of their proficiency in accurately delineating regions of interest (ROIs) (Yuan et al., 2022). Detection and classification jobs frequently employ models such as YOLO, SSD, and the R-CNN series, which are proficient in recognizing and categorizing objects in photos (Vilcapoma et al., 2024). GANs and their derivatives, such as DCGAN, cGAN, CycleGAN, and Pix2Pix, are utilized for data augmentation, picture reconstruction, and style transfer, improving training data’s diversity and quality (Simion et al., 2024). Three-dimensional (3D) CNNs and their variants, including 3D U-Net, are explicitly engineered for volumetric data, which is essential for examining 3D structures in CT scans (Wen et al., 2023). Furthermore, sophisticated models such as Dual-Path Networks and those developed by neural architecture search are reviewed to enhance performance in CT image processing, guaranteeing that the models are efficient and precise (Min et al., 2024).

Deep learning comprises three fundamental methodologies: supervised, semi-supervised, and unsupervised (Liporaci et al., 2024). For medical image analysis tasks, including disease detection and classification, CNNs have grown into the predominant DL architecture (Odimayo et al., 2024). A schematic representation of a typical CNN architecture is shown in Figure 5. Widely utilized during the diagnosis of numerous illnesses, CNNs specialize in extracting pertinent information from medical pictures, including X-rays, CT scans, and MRI scans. While CNNs excel at processing spatial information in medical images, recurrent neural networks (RNNs) and their variants, such as long short-term memory, have proven effective in handling sequential data, which is often encountered in healthcare settings (Chae et al., 2024). Generative adversarial networks (GANs) have emerged as a powerful tool to address the limited availability of labeled data by generating synthetic medical data that can be used to augment the training datasets (Su et al., 2023). Among these techniques, GANs stand out for their unique capabilities in data augmentation and image quality enhancement. GANs consist of two neural networks, a generator and a discriminator, that engage in a competitive training process. While the discriminator assesses their veracity against actual data samples, the generator generates new data instances. The generator learns to create more realistic data because of this adversarial process, which is especially useful in medical imaging where class imbalance and data scarcity are frequent problems. GANs can produce synthetic CT images of PTB that resemble actual TB lesions, increasing the data available for DL model training. This improves the model’s generalization across various patient populations and imaging settings and reduces overfitting. Semi-supervised learning methods, which leverage labeled and unlabeled data, have shown promising results in tasks like medical image segmentation and disease classification (Tseng et al., 2024). These techniques can help improve model performance when the labeled data is limited. Furthermore, unsupervised learning approaches, such as clustering and anomaly detection, have been utilized to identify novel disease subtypes, detect rare diseases, and uncover hidden patterns in medical data (Thabtah et al., 2022). These methodologies can yield essential insights for physicians and scientists, potentially facilitating the identification of novel disease biomarkers and enhanced patient screening.

Conventional techniques for PTB detection frequently depend on radiologists’ expertise, which may be subjective and protracted. The integration of DL into this process has shown promising advancements. The application of DL in TB management is shown in Figure 7. Zhang et al. (2024) introduces a 3D multi-scale attention residual network (MAResNet) to recognize PTB utilizing CT images. MAResNet is the integration of the Convolutional Block Attention Module (CBAM) alongside residual modules. This dual mechanism enhances the distinguishability of image features and allows for the efficient reuse of shallow features. The accuracy of MAResNet in classifying PTB reaches 94%, which is essential for differential diagnosis and treatment planning.

Another study by Yoon et al. (2023) a 3D neural network model, nnU-Net, will be created to investigate the clinical significance of CT cavity volume and evaluate the model’s efficacy in cavity detection. The research retrospectively analyzed 392 patients with mycobacterial pulmonary disease, including TB and non-tuberculous pulmonary disease. The nnU-Net model demonstrated high sensitivity in detecting cavities, with a mean DSC of 78.9. One notable application is the cascading deep supervision U-Net model, as highlighted in the study by Hu et al. (2022), which concentrates on the diagnosis of pneumoconiosis complicated by PTB. The CSNet model leverages the strengths of HRCT to provide high-resolution imaging coupled with the DL framework to enhance the segmentation and diagnosis of affected lung tissues. This approach has shown superior performance over traditional U-Net models, with an AUC value of 0.947.

In addition to identifying PTB, DP is adept at monitoring changes in patients’ conditions following treatment and assessing the severity of the disease.

The accurate quantification of lesion progression in PTB from CT images is pivotal for disease monitoring and treatment response evaluation. The integration of depth information in the ResNet model allowed for capturing the 3D characteristics of pulmonary lesions, providing a more comprehensive analysis than traditional 2D image assessments. This approach underscores the potential of DL in identifying the presence of disease and quantifying its extent and severity. Gao et al. (2020) developed a 3D ResNet incorporating depth information at each layer, the suggested depth-ResNet model demonstrated remarkable performance, with an average classification accuracy of 92.7% in predicting severity scores. Wu et al. (2019) utilize DL to create a diagnostic framework that detects PTB lesions and classifies them into specific types, such as military and tuberculoma. Applying a Noisy-Or Bayesian function to calculate an overall infection probability enhances the diagnostic report with quantitative analysis, offering clinicians a more thorough comprehension of the infection’s scope and characteristics. The approach utilized advanced 3D CNNs to examine CT imaging datasets from 233 patients with active PTB and 501 healthy controls. The recall and precision for identifying PTB patients were 98.7% and 93.7%, respectively. The classification accuracy of PTB was 90.9%.

Studies have shown that the effective radiation dose from a single CT scan can range from a few millisieverts (mSv) to over 8 mSv, which is significantly higher than the typical annual background radiation exposure of around 3 mSv (Sharma and Surani, 2020; McKenna and McMonagle, 2024). Using ionizing radiation in CT scans has raised concerns about the potential health risks. The mean radiation exposure for ultra-low-dose computed tomography (ULDCT) ranges from 0.05 to 0.26 mSv, representing a significant reduction when compared to the radiation levels associated with standard-dose CT scans (Heltborg et al., 2024). Yan et al. (2021a) demonstrates the application of a CycleGAN model for denoising ultralow-dose CT images in evaluating PTB. The optimized CycleGAN model improved the peak signal-to-noise ratio by 2.0 dB and the structural similarity index by 0.21, providing satisfactory image quality with lower noise levels than hybrid and model-based iterative reconstruction techniques. The optimized CycleGAN technology might enable chest ULDCT to generate diagnostically acceptable images for TB evaluation.

While the initial validation results are impressive, further prospective validation studies are necessary within actual clinical settings. Once these commercial AI systems have been thoroughly tested, they could offer physicians globally convenient, efficient, and precise diagnostic tools, thereby aiding clinical decision-making in the foreseeable future. Integrating clinical decision support systems is a key challenge in applying DL to CT-based PTB diagnosis. This requires seamlessly integrating AI models with existing healthcare systems to ensure that diagnostic results are effectively communicated to physicians and influence clinical decision-making. Specifically, it requires addressing several challenges: First, integrating the output of AI models with data from existing systems like electronic medical records and PACS to ensure diagnostic results are associated with the patient’s other clinical information. This requires addressing issues around data formats and security. Second, ensuring AI-assisted diagnosis can seamlessly integrate into the physician’s clinical workflow without adding extra steps or disrupting the normal diagnostic process. This requires optimizing and redesigning existing workflows. Finally, physicians must understand the rationale and logic behind the AI model’s diagnoses to evaluate the results and perform secondary confirmations. This requires improving the interpretability of the AI model so that physicians can gain insights into its inner workings. Interpretability remains a critical challenge in the domain of CT-based pneumonia and PTB diagnosis, as DL models used for image analysis are often regarded as “black boxes” due to their high-dimensional and non-linear nature. Providing clear and actionable explanations can help physicians and DL models jointly improve diagnostic accuracy, reducing the risk of misdiagnosis of PTB. One effective approach involves using Grad-CAM to visualize the regions of the lungs that the model focuses on during the diagnosis of PTB, such as lesions or areas of consolidation (Jegatheeswaran et al., 2024). Additionally, Shapley additive explanations or local interpretable model-agnostic explanations can be applied to identify the most significant features contributing to the model’s predictions, such as pixel intensities or specific ROIs (Chung et al., 2024; Peng et al., 2024). Interactive tools like heatmaps can further enhance interpretability by allowing physicians to explore the model’s behavior and examine specific predictions in detail (Liu et al., 2020). It is also important to evaluate interpretability methods with physicians to ensure that the explanations are comprehensible and clinically relevant. Finally, ensuring compliance with regulatory standards and incorporating feedback from medical professionals will further support the safe and effective deployment of these systems in actual clinical settings.