Application of Artificial Intelligence in Thoracic Radiology: A Narrative Review
Article information
Abstract
Thoracic radiology has emerged as a primary field in which artificial intelligence (AI) is extensively researched. Recent advancements highlight the potential to enhance radiologists’ performance through AI. AI aids in detecting and classifying abnormalities, and in quantifying both normal and abnormal anatomical structures. Additionally, it facilitates prognostication by leveraging these quantitative values. This review article will discuss the recent achievements of AI in thoracic radiology, focusing primarily on deep learning, and explore the current limitations and future directions of this cutting-edge technique.
Introduction
Recent advancements in artificial intelligence (AI), particularly in deep learning (DL), have profoundly influenced the medical field over the past decade, transforming the analysis and interpretation of radiological images. Radiology, especially thoracic radiology, has been at the forefront of adopting AI, owing to its inherently digital workflow and the availability of substantial, standardized datasets [1]. The analysis of X-ray-based images, such as chest radiography (CXR) and computed tomography (CT), is critical in thoracic radiology. Detecting and interpreting subtle differences in lesion density, a task where AI has demonstrated considerable promise, is essential for identifying abnormalities. Studies demonstrate that AI’s performance in analyzing CXR images matches or surpasses that of thoracic radiologists [2-9], prompting its exploration in various clinical scenarios [10-17]. More recently, attention has shifted towards the potential of AI, particularly DL, in enhancing chest CT image analysis [18-22].
This review article aims to shed light on the current achievements and applications of AI, with a specific focus on DL, in thoracic radiology. We will examine how DL has been implemented across various imaging modalities (i.e., CXR and CT) and in the diagnosis and management of major pulmonary diseases, such as lung cancer, chronic obstructive pulmonary disease (COPD), interstitial lung disease (ILD), and pulmonary embolism. Additionally, we will address the limitations of current AI technologies, including their dependence on high-quality data and challenges related to generalizability. Future directions for AI application in clinical practice, including an overview of generative AI and the necessity for integration with existing healthcare systems, will also be discussed.
Applications of AI in CXRs
Initial applications of AI predominantly focused on detecting radiologic abnormalities within CXRs. AI is capable of identifying a wide range of radiologic abnormalities, such as pneumothorax, pneumoperitoneum, mediastinal widening, nodule or mass, consolidation, pleural effusion, atelectasis, fibrosis, calcification, cardiomegaly, and interstitial opacities [10,11,23]. This section will review previous studies specifically addressing the detection of consolidation, pneumothorax, and nodules using AI; it includes a Table 1 summarizing examples of AI applications in CXRs.
Application of artificial intelligence in thoracic radiology
1. AI application for pulmonary infection
In patients with pneumonia, the airspaces fill with pathogens and inflammatory cells, leading to the appearance of ground-glass opacity and consolidation, the most common radiological signs of this condition [24]. CXR remains a traditional modality for diagnosing pneumonia. Previous studies have shown that DL can identify pneumonia in CXRs with high accuracy (sensitivity, 82.4%–88.7%; specificity, 72.8%–78.2%), and significantly, it has outperformed thoracic radiologists in detecting consolidation [11,23]. One meta-analysis documented that DL accurately distinguished patients with pneumonia from those without, achieving a pooled sensitivity of 98.0% and a pooled specificity of 94.0% [25]. Moreover, DL has shown capability in differentiating causative pathogens (i.e., bacterial vs. viral) with a pooled sensitivity and specificity of 89.0% each [25].
DL can be effectively utilized in specific situations. Utilizing DL, radiologists have detected pneumonia more sensitively in patients with febrile neutropenia (DL vs. radiologist-alone: pooled sensitivity, 79.4% vs. 75.2%), while maintaining comparable specificity (pooled specificity, 76.8% vs. 75.4%) [8]. Additionally, emergency department physicians can benefit from DL in making accurate diagnoses when evaluating patients with acute febrile respiratory illness and detectable lesions in CXRs (DL vs. physician-alone: pooled sensitivity, 82.2% vs. 73.9%; pooled specificity, 98.1% vs. 88.7%) [26].
2. AI application for pulmonary tuberculosis
Pulmonary tuberculosis is another area where DL-assisted detection and monitoring of treatment efficacy could be beneficial. DL-based detection algorithms have demonstrated superiority over physicians in identifying active pulmonary tuberculosis, with area under the curves (AUCs) of 0.99 for classification compared to 0.75 for non-radiology physicians, 0.95 for board-certified radiologists, and 0.97 for thoracic radiologists. For localization, the AUCs are 0.99 versus 0.66 for non-radiology physicians, 0.90 for board-certified radiologists, and 0.93 for thoracic radiologists [5]. Qin et al. [27] evaluated five commercially available algorithms in high-prevalence settings and found that their performance in detecting active pulmonary tuberculosis surpassed that of human readers, reducing the required number of Xpert MTB/RIF assays by 50% without compromising sensitivity. DL-based algorithms have also shown efficacy in low-prevalence mass screening settings for detecting active pulmonary tuberculosis [4,14]. Further, a DL-based algorithm has shown promise in assessing disease activity, with performance comparable to that of radiologists [9].
Among patients with positive interferon-gamma release assay results, a DL-based algorithm for identifying active pulmonary tuberculosis showed higher sensitivity compared to radiologists at a fixed specificity (72.7% vs. 59.5% at a specificity of 88.0%) [12]. The sensitivity of radiologists improved with DL-assisted interpretation (72.3% vs. 59.5%) [12]. In a hypothetical scenario where DL-based prescreening was implemented prior to radiologist interpretation, specificity exceeded that observed with radiologists alone (88.8% vs. 85.7%), while the workload was reduced by 85.2% (from 1,780 cases to 263) [12]. This indicates that DL could function as a second reader to enhance sensitivity or as a prescreener to reduce the workload of radiologists.
3. AI application for pulmonary nodule
Identifying pulmonary nodules in CXRs is crucial, as this radiologic sign is a prevalent indicator of lung cancer. Consequently, nodule detection in CXRs has attracted significant interest in early studies, and the deployment of DL has generally enhanced the detection capabilities of physicians, including thoracic radiologists (Figure 1) [6,7,11,23,28]. Hwang et al. [28] demonstrated that an AI-aided detection system improved diagnostic yields for newly visible metastases in cancer patients undergoing CXRs (0.86% vs. 0.32%), without significantly increasing false-referral rates (0.34% vs. 0.25%). In a single-center randomized trial, the detection rates of actionable lung nodules (category 4 according to Lung Imaging Reporting and Data System) and malignant lung nodules in a CXR-based health screening population were higher with the support of DL-based software (actionable lung nodules, 0.59% vs. 0.25%; malignant lung nodules, 0.15% vs. 0%) [29]. Notably, this study also found that false-referral rates were comparable between the DL and non-AI groups (45.9% vs. 56.0%) [29]. Jang et al. [30] reported that nodules were initially missed in a significant number of patients with lung cancer on their CXRs, and that DL could reduce these false-negative reports. Although the definitions of some key outcomes, such as false-referral rates, varied across studies [28-30], current evidence suggests that implementing DL in CXR interpretation may improve the detection rates of clinically significant pulmonary nodules, without increasing false-referral rates. From this perspective, the results of the ongoing interventional randomized controlled trial (LungIMPACT) are eagerly anticipated to determine whether AI implementation could shorten the time to diagnose patients with suspicious findings for lung cancer [31].
Example of deep learning-based automated detection for pulmonary nodules (A, B) and pneumothorax (C, D) using the commercially available algorithm (INSIGHT CXR v3.1.4.1, Lunit).
4. AI application for pneumothorax
Theoretically, pneumothorax can be easily identified in CXRs by recognizing the visceral pleural edge and air density [24]. However, in clinical practice, small pneumothoraces may sometimes be overlooked. Conversely, DL demonstrates high accuracy in detecting pneumothorax (Figure 1) [11]. Thus, using DL as a secondary reader could significantly reduce false-negative rates in detecting pneumothorax [3,7,10,11,15,32]. Specifically, DL-assisted detection of pneumothorax is promising for use in emergency departments and in monitoring delayed-onset pneumothorax following percutaneous transthoracic needle biopsies [3,15,32,33]. Furthermore, since DL can estimate the volume of pneumothorax in CXRs [34], its implementation could improve the management and monitoring of patients with pneumothorax.
5. Other applications of AI in CXRs: beyond the detection of a single abnormality
In practice, physicians do not evaluate CXRs in a task-specific manner, as these images may contain multiple abnormalities simultaneously. Hwang et al. [7] reported that a DL-based detection algorithm enhanced performance in both image-wise classification and lesion-wise localization for physicians assessing CXRs with multiple abnormalities (AUCs: for image-wise classification, 0.81–0.93 to 0.90–0.96; for lesion-wise localization, 0.78–0.91 to 0.87–0.94). Kang et al. [35] conducted a similar study using CT scans as the ground truth. Moreover, other studies have shown that thoracic radiologists could benefit from DL-assisted detection tools, which also reduce their interpretation times for CXRs [10,11,23].
In routine clinical practice, identifying interval changes in CXR findings, such as pulmonary opacities, pneumothorax, or effusion, in follow-up images is essential. Singh et al. [36] evaluated the potential of DL to identify interval changes in follow-up CXRs. In their study, the DL algorithm had to detect and segment the abnormalities before assessing the interval changes [36]. This approach limited its clinical applicability as it was restricted to specific abnormalities for which it was trained. In contrast, Yun et al. [17] recently developed a DL algorithm using a subtraction-based approach, as opposed to a detection-based one, to identify CXRs without interval changes. This algorithm is expected to reduce the workload for radiologists.
After the introduction of DL, the role of CXRs in patient care has expanded beyond mere lesion detection. Several studies suggest that CXRs could be instrumental in selecting candidates for cancer screening and prognostication. Lee et al. [37] observed that utilizing CXR-based DL was advantageous for choosing appropriate candidates for lung cancer CT screening. This method enhanced the 2021 U.S. Preventive Services Task Force recommendations by reducing the percentage of CT screening candidates (35.8% vs. 45.1%) without affecting cancer detection rates [37]. Additionally, using CXR to estimate biological age through DL could provide crucial prognostic information [38-40]. The CXR-derived biological age strongly correlates with chronological age (correlation coefficient, 0.95 to 0.97) [39,40], and serves as a risk factor for all-cause and cardiovascular mortality [38,39]. Disparities between CXR-based and chronological ages are linked with various chronic comorbidities including hypertension, COPD, ILD, and chronic renal failure [40].
Applications of AI in CT Scans
Table 1 summarizes examples of AI applications in CT scans.
1. AI applications for lung cancer
1) Nodule detection, measurement, and characterization
After the reduction in lung cancer mortality through the use of low-dose CT scans as screening tools [41], the assessment of pulmonary nodules in high-risk individuals using low-dose CT has consistently expanded over the last decade. In parallel with its use in CXRs, DL has shown performance comparable or superior to that of radiologists in detecting pulmonary nodules [42-44]. With the support of DL, radiologists have not only enhanced their detection accuracy but also reduced interpretation time (Figure 2) [45].
Example of deep learning-based automated detection, characterization, and measurement of pulmonary nodules using the commercially available algorithm (LuCAS, Monitor Corporation). Deep learning-based algorithms for detecting, segmenting, and categorizing pulmonary nodules are applicable in both lung cancer screening and non-screening settings, such as metastasis work-ups. These algorithms are capable of measuring the size of nodules, estimating growth rates, and assessing treatment responses in real clinical settings. GGN: ground-glass nodule.
In CT images, it is well-established that ground-glass portions typically correlate with the pathological lepidic components of lung adenocarcinomas, while solid portions are associated with invasive components [46]. Thus, precise quantification of solid and ground-glass portions is critical for managing lung adenocarcinomas that appear as subsolid nodules [47,48]. DL has attained performance levels comparable to those of radiologists in quantifying solid portions of lung adenocarcinomas and distinguishing invasive types from others [20,49,50].
DL can also be used for classifying nodules (benign vs. malignant) and estimating their growth rates [42-44]. For example, DL has enhanced the accuracy of predicting lung cancer risk in lung cancer screening settings [51]. Similar to human readers, the availability of prior CT images can improve the accuracy of malignancy risk estimation by DL models for pulmonary nodules [52].
2) Prognostication in patients with lung cancer
DL-based prediction of histopathologic findings from CT scans is an expanding area of interest in lung cancer research. Predicting visceral pleural invasion in peripheral lung cancer is crucial for planning surgical treatment, although it continues to challenge even expert thoracic radiologists and thoracic surgeons. Previous studies have indicated that DL-based predictions of visceral pleural invasion can achieve performance comparable to that of thoracic radiologists [53,54]. One benefit of DL-based prediction is the capability to tailor sensitivity and specificity based on clinical requirements [54]. Furthermore, DL-based models have successfully predicted high-risk histopathologic features (such as mediastinal nodal metastasis and the presence of lymphatic, venous, perineural, or visceral pleural invasion) in CT images [18,55]. DL models have also shown promise in directly predicting disease-free and overall survival in patients with early-stage lung adenocarcinomas using chest CT scans [19,56-58]. Remarkably, a DL-based prognostic model utilizing CT scans exhibited excellent inter-scan and inter-reader reproducibility, with a Pearson’s correlation coefficient of 0.98 for both [56].
2. AI application for COPD
Density-based metrics (such as the percentage of low-attenuated area below –950 hounsfield unit [HU] and the 15th percentile lung density) and lung volume assessment have been utilized for the quantitative analysis of CT scans in COPD patients [59]. Alternatively, visual grading of centrilobular and paraseptal emphysema has been proposed and validated [60].
However, visual grading of centrilobular emphysema is subject to inter-reader and inter-rater variability, which poses a challenge for its practical application [61]. DL offers a promising solution to reduce this variability. Humphries et al. [62] employed DL to classify the pattern of centrilobular emphysema according to the Fleischner Society criteria. In this study, a moderate agreement between human readers and DL was observed (weighted κ, 0.60) [62]. The use of DL-based classification of centrilobular emphysema improved the predictive accuracy of the models for COPD-associated functional parameters (such as forced expiratory volume in 1 second, the ratio of forced expiratory volume in 1 second to forced vital capacity, 6-minute walk distance, or St George’s Respiratory Questionnaire) [62]. This research also demonstrated that worsening grades of centrilobular emphysema assessed by DL were associated with higher mortality rate [62]. Remarkably, even the presence of trace amounts of centrilobular emphysema detected by DL was identified as a prognostic factor [62]. Additionally, Oh et al. [63] recently reported that increases in the Fleischner Society emphysema grade during follow-ups, as determined by DL-based classification, were associated with higher mortality in the COPDGene cohort.
The precise assessment of centrilobular emphysema appears more significant due to its greater prognostic impact compared to paraseptal emphysema [64]. For example, moderate or greater degrees of centrilobular emphysema, determined using the Fleischner Society criteria, have been linked with accelerated decreases in diffusing capacity and airflow limitation, while paraseptal emphysema has not shown similar associations [64]. Therefore, separately assessing centrilobular and paraseptal emphysema may hold greater clinical relevance.
In this context, a recent study suggested that independent assessments of centrilobular (on a 5-grade scale) and paraseptal emphysema (on a 3-grade scale) using DL enabled more accurate predictions of spirometric obstruction and mortality compared to density-based measures (the percentage of low-attenuated area below –950 HU) [65]. Nonetheless, the model’s performance was suboptimal, achieving accuracies of only 41.7% and 52.8% in stratifying individuals based on centrilobular (5-grade) and paraseptal (3-grade) emphysema severity scores, respectively [65].
Despite recent studies showing the promising potential of DL applications for assessing emphysema [62,63,65], challenges remain. Most evaluations and validations of DL-based centrilobular emphysema grading have been conducted within the COPDGene cohort [66], excluding patients with a history of other lung diseases except asthma. Therefore, the generalizability of DL-based centrilobular emphysema grading in patients with emphysema who also have other lung conditions, such as pulmonary fibrosis or extensive bronchiectasis, needs further investigation. Moreover, the performance of DL-based centrilobular and paraseptal emphysema classification requires enhancement for effective implementation in daily clinical practice.
3. AI application for interstitial lung abnormality and ILD
Interstitial lung abnormality (ILA), a radiological and incidental finding detected on CT scans, has recently drawn attention from physicians [67]. Various radiologic risk factors for ILA progression, such as a predominance of basal and peripheral distribution and the presence of fibrotic features, have been identified [67]. In patients with ILDs, both the extent of the disease and the radiological subtypes are well-established factors that affect prognosis [68,69].
From these perspectives, accurately assessing the extent and categorizing the subtypes of ILAs and ILDs are critical. Although previous studies have reported high inter-reader agreement on the presence of ILAs, agreement on the subtypes of ILAs has varied across studies [70-72]. Additionally, radiologists tend to overestimate the extent of interstitial abnormalities, especially when the extent of ILAs is less than 5% [70]. In this context, DL has demonstrated potential for enhancing inter-reader agreement and predicting the prognosis for patients with ILDs [21,73-76]. For example, DL could reduce inter-reader variability in categorizing fibrotic ILD based on the Fleischner Society criteria (median weighted κ; between radiologists, 0.56 [interquartile range, IQR, 0.55 to 0.58]; between DL algorithm and radiologists, 0.64 [IQR, 0.55 to 0.72]) [76]. A DL-based pattern of usual interstitial pneumonia was found to be prognostic in patients with progressive fibrotic ILD, whereas the usual interstitial pneumonia pattern determined by radiologists showed borderline significance [75]. Additionally, DL models enable physicians to quantify each ILD-associated feature that indicates fibrotic lungs, such as ground-glass opacities, reticular opacities, or honeycombing cysts (Figure 3) [21,70,74].
Example of deep learning-based texture analysis of interstitial lung disease using a commercially available algorithm (AVIEW, Coreline Soft). An automated algorithm for segmenting abnormal findings in volumetric computed tomography scans is available (yellow, ground-glass opacities; orange, reticular opacities; pink, consolidation; red, honeycombing; blue, emphysema). This algorithm is useful for assessing the severity of interstitial lung disease and monitoring the progression of pulmonary fibrosis.
4. AI application for pulmonary embolism
CT pulmonary arteriography provides a timely and highly accurate diagnosis for pulmonary embolism [77]. A meta-analysis reported a pooled sensitivity of 88.0% (95% confidence interval [CI], 80.3% to 92.7%) and a pooled specificity of 86.0% (95% CI, 75.6% to 92.4%) in detecting pulmonary embolism [78]. Liu et al. [79] demonstrated that clot burden can be fully automated through volumetric assessment, which correlates with right ventricular functional parameters. Wiklund and Medson [80] illustrated the effectiveness of DL in detecting cancer-associated incidental pulmonary embolism, with a prevalence increase from 0.8% to 2.5% following DL implementation, and reported significant improvements in turnaround times for reporting and initiating treatment (median report time reduced from 24.68 to 0.66 hour; median treatment time from 28.05 to 0.98 hour). Conversely, another report noted that despite acceptable performance, DL implementation did not reduce the time for reporting and initiating treatment in the emergency department [81].
5. Other applications: assessment of thoracic body composition using chest CT scans
DL-based CT body composition analysis has widely been adopted for opportunistic screening [82,83]. DL models enable automated, multi-level two-dimensional or volumetric segmentation of body compositions, such as skeletal muscles and adipose tissues, and achieve excellent segmentation performance (dice similarity coefficients >0.9) within a short timeframe [84]. Consequently, these body composition measures have provided valuable insights into patient outcomes [82,83,85,86]. For instance, DL-derived thoracic body composition metrics have proven useful in predicting mortality from lung cancer, cardiovascular disease, and all causes among individuals undergoing lung cancer screening [82]. Additionally, DL enables more accessible assessment of thoracic skeletal muscles, such as the pectoralis and erector spinae, which are known prognostic factors in patients with COPD (Figure 4) [87,88].
Example of skeletal muscle segmentation using a deep learning-based automated algorithm (AVIEW, Coreline Soft). Deep learning-based algorithms for segmenting skeletal muscles can be used to assess the quantity (area) and quality (density) of thoracic skeletal muscles in patients with pulmonary diseases (red, pectoralis muscles; blue, erector spinae muscles; orange, intercostal muscles; yellow, serratus anterior and subscapularis muscles; green, dorsal back muscles, excluding erector spinae muscles). Automated quantification of thoracic skeletal muscles is expected to aid in prognosticating patients with major pulmonary diseases like lung cancer, chronic obstructive pulmonary disease, and interstitial lung disease.
Similar to abdominal visceral adipose tissue, associations between the characteristics of epicardial adipose tissue and systemic diseases have been suggested [89,90]. The volume and attenuation of epicardial adipose tissue are significant predictors of future cardiovascular events in asymptomatic individuals [90-92]. With the support of DL, both volume and attenuation of epicardial adipose tissue can be measured accurately in a fully automated manner [92-94]. Notably, automated segmentation of epicardial adipose tissue is achievable from non-gated, low-dose chest CT scans [94]. Given the high cardiovascular mortality rates among individuals undergoing lung cancer CT screening [95], this technique is deemed useful for identifying high-risk screenees.
Limitations and Future Directions of AI Application
There exists an imbalance between the rapid advancement in AI performance within medical imaging and the lag in developing methods for explainability, which presents challenges for clinical trust, patient safety, and regulatory approval. A critical issue is the ‘black-box’ problem, in which high-performing models lack transparency, making it difficult for clinicians to understand or validate diagnoses. Interpretability is crucial in this field, yet existing methods such as gradient-weighted activation maps or saliency maps, which provide image-wise and lesion-wise insights [7], often offer local rather than global explanations, resulting in inconsistent or ambiguous outcomes and complicating clinical application [96].
A recent study highlighted that this form of image-based explanation fell short in clinical practice [97]. The visualizations did not effectively guide physicians in interpreting AI results [97]. Therefore, model interpretability must continue to evolve to better clarify its decision-making process, using both visualization and textual descriptions, especially in the context of large language models. This is particularly relevant for prognostic models, where the target prediction is a future event not visible in the images being analyzed. Alternatively, uncertainty quantification has garnered attention in DL for medical imaging analysis [98]. The inherent uncertainty of DL models may stem from labeling errors or insufficient training data [98]. Therefore, measures of uncertainty, reflecting model trustworthiness, should be included in DL model outputs along with predicted probabilities [98].
The quality and quantity of training datasets are crucial in the development of AI models. AI models trained on datasets biased in terms of demographics, disease severity, imaging hardware, or imaging acquisition parameters may not generalize well across various clinical settings [99]. To achieve a more generalizable AI model, it is ideal to collect data from geographically diverse institutions with different medical volumes. Federated learning presents a potential solution by enabling AI models to be trained across multiple decentralized institutions, thereby helping to create a more generalizable dataset without compromising data privacy [100]. However, even with adequate training datasets, successful training does not always ensure optimal results at each institution. Allowing for institution-specific fine-tuning of AI models can help address performance disparities caused by variations in imaging hardware or acquisition parameters, ensuring that the AI model meets the specific needs of each site [101]. If institution-specific tuning is not feasible, validating the model using the institution’s own dataset should be mandated before deploying AI models in clinical practice.
Robustness in test-retest reproducibility is another crucial factor for implementing AI-based techniques in real-world clinical settings. Specifically, the DL-based analysis of CXRs can be influenced by variables such as the patient’s respiration status, positional changes, and image post-processing techniques. Similarly, variations in density in chest CT images due to changes in acquisition parameters and respiration status may affect image quality. Lower reproducibility could limit the clinical application of these advanced techniques, particularly when assessing treatment response or evaluating pulmonary function decline using CXRs [28,102]. Currently, there are limited studies focused on the reproducibility analysis of AI models. To evaluate reproducibility, one approach is to use a pair of stable CXRs taken within a month [102]. Alternatively, test-retest reproducibility could be assessed using synthetic images [103].
The recent advent of generative AI, including large language models, vision-language models, or foundation models, has captured significant interest in the field of medical imaging [104-106]. These models can generate radiologic reports [106]. However, overcoming ‘hallucination’ in generative AI remains a challenge [105,106]. Moreover, the interaction between humans and AI, particularly in the context of large language models, remains largely unexplored.
Advancements in medical AI have provided evidence that AI could enhance current medical practices. For example, AI can reduce interpretation time and improve reader performance in CXR analysis [10,23]. However, in medical fields, enhanced lesion detection does not necessarily lead to improved patient outcomes [2]. While DL shows promise as a secondary reader in radiology [2], it is essential to recognize that DL could increase false chest CT referrals without positively impacting referral rates and may not consistently reduce the time for management in certain contexts [13,81]. Identifying the ‘sweet spots’ where AI can genuinely benefit patient management is crucial. The application of AI to chest CT images, which contain more data than CXRs, could overwhelm physicians with information, especially with the emergence of generative AI in medical imaging. Therefore, careful deliberation on how to implement AI techniques without disrupting radiologists’ daily workflow and effectively integrating and conveying AI-driven metrics to clinicians to enhance patient care and ultimately improve patient outcomes is vital [2].
Conclusion
In thoracic radiology, AI has demonstrated the potential to enhance the quality of current medical practices. This improvement is achieved by enhancing diagnostic accuracy, predicting patient outcomes, providing clinically relevant information previously difficult to obtain due to time limitations, and reducing the interpretation time. Given the accumulated body of evidence, it is critical to consider how these advancements can be structured to support radiologists’ performance and improve patient outcomes in real-world clinical settings.
Notes
Authors’ Contributions
Conceptualization: Kim H. Methodology: Kim H. Formal analysis: Lim WH. Data curation: Lim WH. Software: Lim WH. Validation: Kim H. Writing - original draft preparation: Lim WH. Writing - review and editing: all authors. Approval of final manuscript: all authors.
Conflicts of Interest
Woo Hyeon Lim received a research grant from Coreline Soft. Hyungjin Kim received research grants from RADISEN and Kakao Brain; consulting fees from RADISEN; holds stock and stock options in Medical IP; holds stock in Soombit.ai.
Funding
No funding to declare.