Artificial intelligence applications in personalizing lung cancer management: state of the art and future perspectives

Introduction

Lung cancer is still a leading cause of cancer-related deaths worldwide. The late stage of diagnosis and the heterogeneity of imaging features and histopathology pose significant challenges for clinicians, in particular, to choose the best treatment option. Non-small cell lung cancer (NSCLC) is the most prevalent subtype of lung cancer; the mainstay of therapy is surgery, chemotherapy, radiation, immunotherapy, or molecularly targeted therapy. Although there has been considerable progress in lung cancer therapy, the overall survival rate continues to be low, and individualized approaches are necessary to enhance patient outcomes (1). It is widely believed by the scientific community that progress in lung cancer therapy depends on the creation of cutting-edge methods that recognize the heterogeneity of the disease enabling personalized treatments for individual patients. In recent years, artificial intelligence (AI) has arisen as a valuable resource in oncology, particularly in the field of image detection and evaluation. The term AI refers to the use of computational technologies to emulate human-like intelligent behavior and analytical reasoning. John McCarthy initially defined AI in 1956 as the science and engineering of constructing intelligent machinery. AI is a field within computer science that includes a series of algorithms able to analyze large-scale big data to execute complex functions, emulating human intellect (2).

Machine learning (ML) is a subset of AI that formulates algorithms primarily dependent on prior, predefined data without direct programming. Deep learning (DL) is a sub-discipline of ML (Figure 1) rooted in a neural network framework modeled after the human brain (Figure 2); DL algorithms do not have to specify features in advance, allowing them to autonomously discover features by exploring the data independently. This data-driven mode makes it more insightful and useful. Today, convolutional neural networks (CNNs) are the most widely used type of DL architecture in the domain of medical image analysis (3).

Figure 1 Implementation of deep learning as a form of supervised learning within the subset of machine learning methods in artificial intelligence (created with BioRender.com).

Figure 2 Analogous parallels between artificial neural networks and biological neural networks. Hidden layers in artificial neural networks can be compared to brain interneurons (created with BioRender.com).

These methodologies are used in imaging in the field of radiomics, where accumulating evidence demonstrates that this can be utilized for quantitative assessment of tumors for activities such as disease characterization or predicting outcomes. The potential of these approaches is huge, and it has been revealed in helping clinical experts to uncover cancer characteristics that fail to be appreciated by naked eyes (4), significantly surpassing the human operator and all earlier related technologies in image recognition and analysis. In this context, radiomics is a nascent and rapidly progressing discipline that merges insights from radiology, oncology, and informatics, highlighting the convergence of medicine and engineering (5). Growing evidence suggests that radiomics can be utilized for quantitative analysis of tumors for activities such as disease characterization or outcome prediction, marking a significant research trajectory in medical applications (6).

Materials and methods

This narrative review is based on a selective literature search carried out in PubMed and Cochrane Library databases from origin to December 2022, with the aim of finding relevant studies using combinations of the following search terms: “Artificial Intelligence (AI), Radiomics, Deep Learning (DL), Machine Learning (ML), lung cancer, lung malignancy, lung nodules, Non-Small Cell Lung Cancer (NSCLC)”. The search strings on PubMed and Cochrane Library were (“AI and Lung nodules”), (“Lung cancer and Artificial Intelligence”), and (“NSCLC and AI”). Text (Free Full texts), Article type (All except Books and Documents), (Humans [Filter]), and (English [Filter]) without any limits on the publication year. These publications were then selected by two distinct authors (F.L. and G.G.). The inclusion/exclusion criteria were reported herein:

• Inclusion criteria were (i) english-language; (ii) article types were: clinical trials, randomized controlled trials, case-control studies, prospective or retrospective cohort studies, review and meta-analyses.
• Exclusion criteria: (i) article not written in English; and (ii) the following kind of articles: books, documents, editorial comments, abstracts, case reports, guidelines, or consensus statements.

Two authors (F.L. and G.G.) independently reviewed the full texts of the identified papers. A third author (M.C.) resolved emerging discrepancies. The selected articles were then examined in full, processed, and summarized to align with the objectives of the review. Finally, the chosen papers were read by all authors and discussed to draw reliable conclusions.

AI and the solitary pulmonary nodule

DL and computer-aided detection (CAD) for distinguishing between benign and malignant tumors

DL solutions are currently leading the field in pulmonary nodule detection. DL algorithms have demonstrated similar or superior performance to medical professionals in various healthcare environments. Their applications include detecting pulmonary nodules in chest radiographs or computed tomography (CT) scans, improving candidate identification for r lung cancer screening (LCS) and forecasting the malignancy of pulmonary nodules (7). Radiologists face hurdles in detecting lung cancer from chest radiographs due to limited contrast resolution. This constraint can obscure lesions that overlap with anatomical structures, leading to higher rates of false-negative diagnoses (8). Lesion size, shape, and location are all independent factors in detection error and can lead to missing lesions during the interpretation of chest radiographs (9). CAD software was developed in the 1970s to increase the accuracy of chest radiography for nodule detection (10). The use of computer algorithms in CAD systems can aid in identifying various medical conditions. This includes the use of computer-aided detection systems (CADe) to help detect irregularities or lesions in medical images, and computer-aided diagnosis systems (CADx) to assist in the analysis and diagnosis of medical images. Such systems have the potential to improve the accuracy and speed of medical diagnosis, particularly in situations where human evaluation may be limited or prone to errors. CAD systems can aid in the detection of chest CT pulmonary nodules by initially screening a large number of chest CT images and highlighting suspicious lesions. This can help radiologists to make secondary discrimination, decrease their workload, and improve screening efficiency. Traditional CADe systems usually consist of two main stages: (I) selection of candidate nodules and (II) removal of false positive nodules (FPN) while preserving true positive nodules (TPN) (11). By enhancing these steps, more cases of lung cancer that may have been missed during routine chest imaging can be identified. This could lead to improved survival rates (12).

The use of low-dose computed tomography (LDCT) scans in LCS has led to a significant increase in the volume of images that radiologists must analyze, which can be a challenging and time-consuming task. Furthermore, the elevated frequency of false positives is an additional concern, as it may lead to superfluous diagnostic evaluations and invasive medical procedures for patients who are not afflicted with lung cancer. These issues highlight the need for more efficient and accurate methods of analyzing LDCT scans for LCS. As known, the low dose CT scan differs in quality to CT scan performed for staging, and for this reason, images acquired with this instrument are lower in quality. This is one of the main reasons for which it is hard to give a definitive definition only on the basis of the images. As shown in Table 1, a comparison among studies collecting the use of CAD or DL has been made. AI using CAD or DL methods has been implemented to decrease radiologist errors and increase the detection rate of pulmonary nodules. However, before employing DL techniques, the CAD system realized only poor detection performance (sensitivity of 70% or lower) for lung nodules and a markedly high false positive rate, which was inadequate for clinical application (13).

The Automated Nodule Detection 2009 (ANODE09) study was the pioneering web-based framework for assessing nodule detection algorithms from LCS CT scans (14). The primary limitations of this study were the dataset size and the uniformity; all the images were acquired from a single center, using the same scanner and protocol. To address these limitations, the Lung Nodule Analysis 2016 (LUNA16) trial was established using 888 scans with 1,186 nodules from the Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI) for training and established (15).

To guarantee robustness, the reference values for each scan were based on annotations from four radiologists. The best algorithm achieved a sensitivity of 97.2%, with an average of one false positive per scan. As a result, it continues to be used as a benchmark for more recent AI algorithms (18).

Christe et al. demonstrated that combining a human observer with any CAD system resulted in higher sensitivity for detecting lung nodules compared to combining two different CAD systems (16). Ciompi et al. developed an AI algorithm to categorize pulmonary nodules into solid, part-solid, non-solid, perifissural, calcified, and speculated employing a dataset of 943 subjects with 1,805 nodules from the Multicenter Italian Lung Detection (MILD) trial (17). They showed that the system’s accuracy was comparable to the inter-observer variability of four experienced human readers, thereby confirming the effectiveness of the system at classifying nodules equivalently to an independent human expert. Radiologists have introduced the concept of “nodule type” in interpreting CT scans, where they must differentiate between various types of opacities based on their appearance and, more importantly, their likelihood of being malignant. In this context, the range of variability reached by the proposed system makes it the first suitable system for automatic analysis of CT scans in LCS. This study indicated that a DL system could assist in reliably categorizing nodule types, which influences the management strategy for LCS participants. According to the National Institute of Health and Care Excellence, for individuals undergoing a chest CT scan as part of targeted LCS, AI-derived CAD software technologies have the potential to be economically beneficial. Therefore, although there is not yet sufficient evidence to endorse the software as standard practice, centers may implement it as part of targeted LCS. Proof needs to be generated to ensure the potential advantages of using the software are achieved in practice and to facilitate comparisons of the various technologies (19).

Malignancy prediction

The existing management strategies for indeterminate pulmonary nodules rely on qualitative or quantitative estimates of the malignancy risk of these nodules. The prominent examples include Lung-RADS by the American College of Radiology, guidelines by the Fleischner Society, and the British Thoracic Society (20). Lung-RADS^® is a quality control mechanism that aims to standardize the reporting and management recommendations for LCS CT. It is designed to minimize confusion in the interpretation of LCS CT results and to streamline outcome monitoring. A comprehensive atlas and lexicon will be established to facilitate this process. The atlas will feature a medical audit and outcome monitoring process, while the lexicon of LCS CT terminology and reporting format will standardize the language used in reports. Many experiences in literature are now reporting the role of AI in LCS (see Table 2).

The Brock model, also known as the Pan-Canadian Early Detection of Lung Cancer or PanCan model, was developed in 2013 to estimate the risk of malignancy in pulmonary nodules (25). The model includes various predictors such as patient demographics, nodule size, type, and morphology. The Brock model is currently integrated into the British Thoracic Society nodule management guidelines (26,27) and is recommended in the Lung-RADS version 1.1. Although the Brock model has demonstrated that radiologists can more precisely evaluate the malignancy risk of a nodule, there is no consensus when radiologists are requested to identify the signs of malignancy (28). In 2020, Baldwin et al. conducted a validation study on lung nodules measuring 5–15 mm, which were identified unexpectedly at three hospitals in the UK. In this study, the lung cancer prediction convolutional neural network (LCP-CNN) was evaluated against the Brock University model, recommended in UK guidelines, and it was found that AI enhances risk prediction (21).

Choi et al. developed a radiomic prediction model for the early detection of lung cancer from lung nodules with low-dose CT in 2018 (22). This model, based on two CT radiomic characteristics, achieved 84.6% accuracy, surpassing the one reported by Lung CT Screening Reporting and Data System (Lung-RADS). The general growth of a nodule on CT and growth not assessed from a single scan only appeared to be the most important predictors of the malignant nature of the nodule. A CNN model was developed by Ardila et al. in 2019 to examine LDCT volumes to detect lung cancer. The model was able to accurately predict the risk of lung cancer over one and two years with an area under the curve (AUC) of 94.4% and 87.3%, respectively (29). The model had greater sensitivity and specificity for lung cancers compared to the human operators when only a single LDCT was available; similar diagnostic accuracy was observed when multiple LDCTs were available (14). Based on this evidence, a superior performance of the DL network methodology was claimed when compared to that of six independent radiologists when dealing with the assessment of malignancy (from one CT scan only).

Although these claims require extensive validation, AI algorithms may have achieved radiologist-level performance for detecting the malignant nature of a pulmonary nodule by CT. Nevertheless, these studies only evaluated individual performances and did not take into account any collaboration between human and machine (30).

Certain tasks that are more challenging for radiologists might be more manageable for the algorithm, and vice versa. For example, subsolid nodules are frequently overlooked by radiologists due to their lower contrast with lung tissue. Conversely, highly irregular nodules might not be identified by the AI because of their rarity in the training dataset (31,32). Regarding the capacity of a human reader to receive assistance from an AI system, two approaches have been outlined: second reader and concurrent reader. When employing computer-aided diagnosis (CAD) as a second reader, the radiologist initially reviews the study independently, submits it to the CAD system, and subsequently re-evaluates the study by concentrating on the CAD marks to finalize the results. Conversely, with CAD as a concurrent reader, the initial independent review by the radiologist is skipped; the study is processed by CAD and then presented to the radiologist, who either accepts or rejects the CAD marks and conducts a final search for any missed nodules (33). Silva et al. showed that CAD provides additional information to radiologists for detecting subsolid nodules in volumetric LDCT for LCS. However, CAD marks necessitate visual confirmation to correct false positives. Integrating CAD with visual reading results in optimal detection performance of subsolid nodules (32). For this reason, this setting would lead to the conclusion that humans and computers would likely work synergistically, leading to better performance in lung cancer diagnosis.

Deep learning-based automatic detection (DLAD) algorithms could precisely identify malignant pulmonary nodules on chest radiographs, occasionally outperforming physicians and improving physicians’ performance when utilized as a “second reader”. In 2018, Nam et al. developed a DL algorithm by using 43,292 chest radiographs. DLAD demonstrated high specificity and successfully identified 100% of high conspicuity nodules, mostly large (>3 cm), nodules, and more nodules in overlap areas than four groups of physicians (8). They reported that the DL algorithm achieved a sensitivity of 71–91%, a specificity of 93–100%, and an AUC of 0.92–0.99, which were superior to the performance of most physicians involved in the study (8).

Compared with previously reported conventional image processing-based computer-aided diagnosis, DLAD exhibited a significantly reduced rate of false-positive findings and delivered high specificity while maintaining sensitivity, resulting in overall better detection performance than thoracic radiologists (34).

In 2020, Ohno et al. (23) analyzed the volumetric change and doubling time of pulmonary nodules of 170 patients with 290 nodules at chest CT, assisted by the CNN applied to computer-aided detection of volume (CADv) measurements. They indicated that the AUC of total volume change per day calculated by the CADv with CNN (AUC =0.94) was notably higher than CADv not using CNN (AUC =0.69), concluding that CNN is a useful tool for improving accuracy and nodule differentiation.

In 2022, Zhang et al. compared the performance of detecting lung nodules between Radiologist Observation and AI-assisted reading. For solid nodules, the accuracy and sensitivity of radiologists were 86.2% and 52.4%, respectively, much lower than 99.1% and 98.8% detected by AI. For partly solid nodules, the sensitivity of radiologists was 23.1%; much lower than the 100% achieved by AI. For non-solid nodules, the sensitivity of human observation was 25.2%, compared to 99.1% of AI. Therefore, AI-assisted reading greatly increased the detection sensitivity of part-solid and non-solid nodules by 74% compared with the radiologists’ observation (24).

AI repeatedly demonstrated itself as a promising innovation for the malignancy prediction of lung nodules. Almost all the studies concluded that incorporating AI into standard radiological diagnostic pipelines will foster enhanced patient care through earlier and more accurate detection of the disease, thereby paving the way toward better outcomes (35). Promising studies are now evaluating the role of AI in the determination of lung nodules malignancy. For example, the DOLCE study is a prospective, observational multicenter study to assess the clinical utility of an AI-assisted CT-based lung cancer prediction tool (LCP) for managing incidental solid and part solid pulmonary nodule vs. standard care (36). Another interesting study is the DART study (also collecting data through the Lung Health Check program) that has the aim of developing new ways of using computer technology (AI) to improve lung health care (37).

AI for predicting histology and other pathological features

Lung cancer is a complex and heterogeneous disease that can be categorized into two main histological types: NSCLC and small cell lung cancer (SCLC). NSCLC represents approximately 85% of all lung cancer cases, while SCLC represents the remaining 15%. NSCLC can be further classified into three subtypes based on the histological features: adenocarcinoma (ADC), squamous cell carcinoma (SCC), and large cell carcinoma. ADC is the most common subtype of NSCLC, accounting for about 40% of all cases, followed by SCC, which represents around 30% of NSCLC cases. Large cell carcinoma is a less common subtype, accounting for approximately 10–15% of NSCLC cases. Each subtype of NSCLC has distinct molecular and genetic characteristics, which may affect the response to treatment and prognosis of the disease. Therefore, accurate subtyping of NSCLC is crucial for personalized treatment decisions and clinical management. Lung biopsy remains up until now the cornerstone for diagnosis of lung cancer, obtained either by image-guided transthoracic needle biopsy or transbronchial by bronchoscopic techniques. Although the histopathologic analysis by experienced pathologists remains the gold standard for diagnosing NSCLC histologic subtypes, it is occasionally challenging to accurately differentiate poorly differentiated ADC and SCC due to similar morphologic characteristics (38) thus requiring confirmatory immunohistochemistry (IHC). Furthermore, it is time-consuming and difficult for a pathologist to analyze highly intricate pathologic images through morphological assessment of tissue sections. In contrast, highly sensitive and automatic AI requires limited human intervention to classify lung cancer histological types/subtypes by analyzing radiological images (CT/PET) and digital histopathological slides using DL algorithms especially CNN (39) to assist (not replace) the pathologists in the decision-making process, making more precise diagnosis of NSCLC (39). Therefore, several studies investigated the various AI models for classifying the histology of NSCLC based on digital histopathological slides and gene profiles, comparing with the performance level of experienced pathologists. Recent studies using AI in pathology aimed to (I) forecast routine diagnostic characteristics (e.g., disease vs. normal tissue, determine tumor grade, and differentiate cancer types) and (II) uncover new insights into disease (e.g., predict gene mutation status, disease recurrence, and outcome) (40). These algorithms can significantly aid pathologists in determining the ratio and distribution of histologic subtypes/patterns in a specimen, which is currently a tedious process (41).

Besides tissue biomarker analysis, liquid biopsies have grown increasingly widespread. Liquid biopsy involves the examination of any tumor-derived product in the bloodstream or serum. Unlike tissue-based biopsies, which evaluate the spatial diversity of the tumor based on the sample site, liquid biopsies can capture temporal heterogeneity by collecting specimens from the bloodstream at various times, potentially revealing the progression of the disease. This technique is useful in various contexts. In early-stage disease, plasma next-generation sequencing (NGS) with comprehensive panels can offer high sensitivity and specificity, aiding in distinguishing benign from malignant nodules, guiding neoadjuvant therapy, and directing adjuvant therapy for patients undergoing surgical resection (42). Circulating miRNA serves as a potential clinical marker for detecting tumors and monitoring the progression of tumorigenesis through liquid biopsy (43). Zhang et al. (43,44) presented a novel computational approach for identifying significant circulating miRNAs that may be applied to early screening, diagnosis, and constant monitoring of lung cancer progression. Specifically, they used synthetic minority oversampling (SMOTE), a technique to assist AI in scenarios where the minority class is underrepresented, risking the overshadowing of its key features by the dominant class. Combined with random forests, this approach was used to detect lung cancer via circulating microRNA (miRNA), attaining an AUC of 0.99. Notably, this study employed a case-control design with samples not limited to early-stage disease, likely inflating performance results. Nonetheless, further exploration into AI/ML for liquid biopsy analysis is essential, as human assessment of such complex, high-dimensional data is impractical. AI is crucial for improving the sensitivity and specificity of liquid biopsies, ultimately benefiting patients through less invasive and earlier cancer detection.

AI models using radiological images for histology classification

Radiologists and oncologists may face challenges in accurately determining the histology of NSCLC tumors due to limited access to invasive techniques for tumor sampling during biopsies. Nevertheless, non-invasive approaches, including the utilization of medical image biomarkers and ML, have the potential to function as effective substitutes for predicting NSCLC histology (Figure 3). In 2020, Pang et al. (44) proposed a DL model to classify SCLC, ADC, and SCC from 2,219 CT images. They found that the DL model was enough powerful to achieve an overall accuracy of 87% and a sensitivity of 90%, 86%, and 92% in classifying SCLC, ADC, and SCC, respectively. Similarly, in the same year, Moitra and Mandal (45) evaluated a DL model on CT scans belonging to 211 NSCLC patients (285,411 CT images). Their study included not only ADC and SCC but also a third non-specified NSCLC histological subtype. Their model reached an overall accuracy of 96.3% and an AUC of 98.5%. While dropout layers were used during training to minimize overfitting by setting some features to zero, all features were considered during the validation process. This artificially enhanced the model’s accuracy, which reached 94.4%. They concluded that their model would be useful in the automated prognosis of NSCLC, and it would help radiologists and oncologists in the decision-making process.

Figure 3 Integrated use of big data (imaging, clinical and metabolic data) through artificial intelligence for predictive purposes at histological, pathological and genetic state.

Recently in 2021, Guo et al. (46) developed a 3D DL and Radiomics methodology to differentiate lung cancer histological types/subtypes (SCLC, ADC, and SCC) from non-contrast CT images, encompassing 920 patients. Results indicated that the DL models can distinguish SCLC, ADC, and SCC with an overall accuracy of 75% (AUC 79–84%) (see Table 3).

AI models using histopathological slide images for histology classification

In 2018, Khosravi (47) and his colleagues from Weill Cornell Medicine collected 12,139 IHC-stained whole-slide images as well as H&E-stained histopathology images from the Stanford Tissue Microarray Database (TMAD) and The Cancer Genome Atlas (TCGA). They trained all CNNs using the chosen images to distinguish between the two histological subtypes [ADC and SCC]. The results showed that complex CNNs could successfully distinguish ADC and SCC across heterogeneous tissue of the tumor slides with no error. Histologic intertumoral heterogeneity in ADCs is described in two terms: frequent (>80%) minor heterogeneity resulting from various growth patterns, such as lepidic, acinar, papillary, micropapillary, and solid in mixed adenocarcinomas (mADCs), and unusual major heterogeneity as seen in the adenosquamous lung carcinomas (AdSqLCs) (38). These tumor growth patterns impact clinical prognosis, as micropapillary and solid patterns are linked to a poorer prognosis (48). Additionally, it is occasionally challenging to distinguish the predominant and minor histologic subtypes; therefore, AI algorithms are applied in different studies to accurately classify ADC growth patterns. In 2019, Wei et al. (49) showed that AI can be used to aid pathologists in classifying the five growth patterns with 90% accuracy thus achieving a pathologist-level performance and ultimately contributing to more accurate grading of lung ADC (see Table 4). Moreover, this study compared CNN models and pathologists with diverse experience backgrounds, the CNN model accomplished slightly improved performance and was superior to inexperienced pathologists.

“Radio-genomics”: AI models for predicting gene mutation in lung cancer

Somatic mutation profiling in oncology has revolutionized clinical practice. The discovery of driver mutations in NSCLC is a prime example. Molecular testing of advanced-stage lung ADC is now the standard of care and a key component of the diagnostic process. For lung ADC, all patients with advanced stages should be tested for epidermal growth factor receptor (EGFR) mutations, anaplastic lymphoma kinase (ALK) and c-ros oncogene 1 (ROS1) rearrangements, and PD-L1 levels to determine their potential response to EGFR, ALK, or ROS1 inhibitors, or immunotherapy, respectively. Additionally, various other biomarkers, including BRAF, ERBB2, MET splice variants and amplifications, and rearranged during transfection (RET) rearrangements, are under clinical evaluation as indicators of response to targeted therapies (50).

In particular, the mutation of EGFR or ALK rearrangements were detected in 15% and 2% of NSCLC patients, respectively (51,52). Currently, the identification of these gene mutations necessitates invasive methods of tissue sampling either by image-guided or transbronchial needle biopsy (TBNB) from the primary tumor or its metastases. To overcome these invasive procedures, liquid biopsy and now AI approaches are trying to investigate/identify tumor mutations to avoid more aggressive procedures (and potentially to reduce the delay for treatment). The combination of medical imaging information derived from radiomics and tumor genomic data is frequently termed “radio-genomics” (53).

Coudray et al. (54) investigated 1,634 whole slide images (1,176 tumor tissue and 459 normal) using the CNN for distinguishing lung ADC, SCC, and normal lung tissue. The DL model demonstrated high performance (AUC =0.97) on par with pathologists. Beside diagnosing lung cancer, Radiomics has been extended to the prediction capability towards the genes’ mutations; the authors have shown that the CNN model was capable of predicting the mutation of six genes (EGFR, KRAS, FAT1, TP53, SETBP1, and STK11) in lung ADC with (AUC =0.73–0.85) based on both frozen sections and HE stained formalin-fixed paraffin-embedded (FFPE) slides. The EGFR tyrosine kinase inhibitor (TKI) targeted therapy is the effective first-line treatment for NSCLC patients with EGFR mutations and offers longer progression-free survival (PFS) and improved quality of life (QoL) compared with chemotherapy. Therefore, many studies investigated the role of Radiomics in integrating CT images for predicting the EGFR mutation. Combining structural and functional imaging with clinical information and the outcome of liquid biopsy allows for the development of models that are most successful at predicting gene mutation. Zhang et al. (55) developed a more complex model using radiomics extracted from the CT images combined with clinical characteristics as (smoking history, sex, and histologic type), for potential discrimination of the EGFR mutation status and showed the ROC (AUC =0.86), which is greater compared to research using only the CT images with the ROC (AUC =0.77) (56).

On the contrary, only a few studies investigated other gene mutations (ALK, KRAS, BRAF, ROS1), and even fewer for the programmed-death-1 (PD-1)/programmed-death ligand (PD-L1) expression. Two Korean studies showed promising results, the first analyzed 172 patients with NSCLC from three hospitals in 2014 (57). They concluded that ALK+ NSCLC has unique features at CT imaging that, when integrated with clinical variables, discriminate ALK+ from non-ALK tumors; it could therefore identify patients with a response to crizotinib with a sensitivity of 83.3%, specificity of 77.9%, and accuracy of 78.8%. The second study, published in 2015 (58), analyzed 539 pathologically proven lung ADC for prediction of ALK, ROS1, or RET by using radiomics from the CT and PET images; it concluded that ALK/ROS1/RET fusion-positive lung ADC exhibits specific clinical and imaging characteristics, allowing effective differentiation of fusion-positive from fusion-negative lung ADC. PD-L1 expression level plays an important role in guiding immunotherapy for NSCLC patients according to the latest version of NCCN (Version 3. 2019). Jiang et al. in 2020 (59) developed a radiomics model based on PET/CT images of 399 stage I–IV NSCLC patients and concluded that radiomics may anticipate PD-L1 expression status in NSCLC relatively accurately with an AUC =0.97 and 0.91 predicting PD-L1 expression rates over 1% and over 50%, respectively (see Table 5).

Limitations of AI-based models

DLAD, although it has demonstrated encouraging outcomes in medical image analysis, faces some constraints because of the requirement for annotated data provided by radiologists, which could mirror the limitations of human perception and analytical judgment. The primary obstacle in applying deep CNNs to medical images is the image quality (60). Although AI models have demonstrated performance comparable to or surpassing that of humans, the intricacy of these models makes them challenging to decipher and comprehend how they arrive at their outcomes, which has resulted in the notion of AI models as “black boxes” (61). Another major concern is the applicability of these models to all patients, which could be resolved by creating continuous learning systems that employ cloud-based methods to enable the real-time delivery of clinical records and ongoing adjustment of the underlying training models. This would guarantee machine-independent consistency of the models (62).

In radiomics, the reproducibility of features is vital due to the variability in image acquisition, preprocessing, and segmentation. PET/CT scans provide extensive imaging data and parametric information but are prone to pitfalls and artifacts and are more costly and technically complex. The lack of a universal consensus on the optimal threshold for lung cancer radiomics is another challenge. Additionally, AI-based approaches require specialized skills, necessitating training for the next generation of radiologists and pathologists. Various AI algorithms have been created in NSCLC patient-focused studies, most of which are retrospective, single-center, with small training samples and lacking external validation, limiting their interpretability and generalizability. Moreover, a well-defined ethical and legal structure from stakeholders (healthcare providers, research organizations, patient advocacy groups, and government) is needed (63).

Conclusions

The application of AI to enhance the understanding and treatment of NSCLC is expanding but its introduction in clinical practice still faces obstacles. However, we may indeed bet that AI-based science will be ubiquitous and indispensable in lung cancer patients. This positive prevision relies on the fact that with the advent of omics sciences and precision medicine, the information obtained from the recalled omics technologies provides physicians with an enormous amount of data (“big data”). Only with the application of AI techniques, researchers and physicians will have the potential to deal with the complexity represented by the quantitative aspect of the big data-related features. For this reason, oncologists, radiologists and surgeons should persist in incorporating machine-learning tools into the clinical management of NSCLC and join the digital transformation that has already occurred in the business and technology sectors. Furthermore, future AI applications for precision medicine in NSCLC could combine radiomics and liquid biopsies into innovative companion diagnostics, offering valuable insights into tumor biology, disease progression, and treatment response in a minimally invasive and longitudinal manner (63).

In particular, as summarized in the present narrative and comprehensive review, there are several possible applications of AI in lung cancer management, with preliminary but very encouraging results:

• Computer-assisted diagnosis will facilitate the identification of early-stage disease, the detection of malignancy in lung lesions, as well as the prediction of histology and other pathological characteristics.
• Advanced AI methods will be useful to predict NSCLC gene mutations without performing a sample biopsy (thus reducing time and costs). Finally, physicians of the future will have powerful tools to predict response to treatments and the occurrence of a tumor recurrence, resulting in a real personalized strategy of care.

Future research perspectives

Several future perspectives have been suggested in the domain of precision medicine for lung cancer therapy; our Team has recently developed a multi-omics platform of research (LANTERN project) (64) that stands as a beacon of innovation, contributing valuable insights that align seamlessly with the evolving paradigm of personalized and preventive healthcare. As we navigate the complexities of lung cancer, the LANTERN project’s multi-omics prospective research design, involving 600 NSCLC patients, presents an unparalleled opportunity to delve into the intricate interplay of genetic, molecular, and clinical factors. By carefully gathering radiomic, genomic, and metabolomic profiles, along with extensive medical and therapeutic information, the study aims to build digital humanized avatars that actively depict each patient’s unique molecular and clinical landscape (Figure 4). This approach not only aims to enhance individualized therapeutic approaches but also has the potential to transform the future of cancer care by aiding the ongoing advancement propelled by sophisticated ML and AI methods. The LANTERN Study’s commitment to ethical standards and compliance with regulations positions it as a pioneering effort that not only addresses current challenges in lung cancer care but also lays the foundation for future breakthroughs in precision oncology. As we reflect on the current trends and future perspectives in AI-based models, the LANTERN project emerges as a pivotal force in enhancing patient care and outcomes in NSCLC, paving the way for a new era in the battle against this challenging disease.

Figure 4 LANTERN Study overview. The figure illustrates key phases of the LANTERN Study, from ‘NSCLC Patients’ through enrollment, and integrated data collection, to the creation of digital humanized avatars, culminating in practical implementation in the clinical setting. NSCLC, non-small-cell lung cancer; DHA, digital human avatar.

Acknowledgments

We thank Dr. Franziska Michaela Lohmeyer for her precious support in English Editing.

Funding: None.

Footnote

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-24-244/prf

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-24-244/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Alduais Y, Zhang H, Fan F, Chen J, Chen B. Non-small cell lung cancer (NSCLC): A review of risk factors, diagnosis, and treatment. Medicine (Baltimore) 2023;102:e32899. [Crossref] [PubMed]
2. Amisha Malik P. Overview of artificial intelligence in medicine. J Family Med Prim Care 2019;8:2328-31. [Crossref] [PubMed]
3. Miotto R, Wang F, Wang S, et al. Deep learning for healthcare: review, opportunities and challenges. Brief Bioinform 2018;19:1236-46. [Crossref] [PubMed]
4. Liu Z, Wang S, Dong D, et al. The Applications of Radiomics in Precision Diagnosis and Treatment of Oncology: Opportunities and Challenges. Theranostics 2019;9:1303-22. [Crossref] [PubMed]
5. Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14:749-62. [Crossref] [PubMed]
6. Qian Z, Li Y, Wang Y, et al. Differentiation of glioblastoma from solitary brain metastases using radiomic machine-learning classifiers. Cancer Lett 2019;451:128-35. [Crossref] [PubMed]
7. Hwang EJ, Park CM. Clinical Implementation of Deep Learning in Thoracic Radiology: Potential Applications and Challenges. Korean J Radiol 2020;21:511-25. [Crossref] [PubMed]
8. Nam JG, Park S, Hwang EJ, et al. Development and Validation of Deep Learning-based Automatic Detection Algorithm for Malignant Pulmonary Nodules on Chest Radiographs. Radiology 2019;290:218-28. [Crossref] [PubMed]
9. Austin JH, Romney BM, Goldsmith LS. Missed bronchogenic carcinoma: radiographic findings in 27 patients with a potentially resectable lesion evident in retrospect. Radiology 1992;182:115-22. [Crossref] [PubMed]
10. Kakeda S, Moriya J, Sato H, et al. Improved detection of lung nodules on chest radiographs using a commercial computer-aided diagnosis system. AJR Am J Roentgenol 2004;182:505-10. [Crossref] [PubMed]
11. Lee MC, Boroczky L, Sungur-Stasik K, et al. Computer-aided diagnosis of pulmonary nodules using a two-step approach for feature selection and classifier ensemble construction. Artif Intell Med 2010;50:43-53. [Crossref] [PubMed]
12. Quadrelli S, Lyons G, Colt H, et al. Clinical characteristics and prognosis of incidentally detected lung cancers. Int J Surg Oncol 2015;2015:287604. [Crossref] [PubMed]
13. Roos JE, Paik D, Olsen D, et al. Computer-aided detection (CAD) of lung nodules in CT scans: radiologist performance and reading time with incremental CAD assistance. Eur Radiol 2010;20:549-57. [Crossref] [PubMed]
14. van Ginneken B, Armato SG 3rd, de Hoop B, et al. Comparing and combining algorithms for computer-aided detection of pulmonary nodules in computed tomography scans: The ANODE09 study. Med Image Anal 2010;14:707-22. [Crossref] [PubMed]
15. Armato SG 3rd, McLennan G, Bidaut L, et al. The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): a completed reference database of lung nodules on CT scans. Med Phys 2011;38:915-31. [Crossref] [PubMed]
16. Christe A, Leidolt L, Huber A, et al. Lung cancer screening with CT: evaluation of radiologists and different computer assisted detection software (CAD) as first and second readers for lung nodule detection at different dose levels. Eur J Radiol 2013;82:e873-8. [Crossref] [PubMed]
17. Ciompi F, Chung K, van Riel SJ, et al. Towards automatic pulmonary nodule management in lung cancer screening with deep learning. Sci Rep 2017;7:46479. [Crossref] [PubMed]
18. Setio AAA, Traverso A, de Bel T, et al. Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: The LUNA16 challenge. Med Image Anal 2017;42:1-13. [Crossref] [PubMed]
19. NICE. Guidance 1 Recommendations. AI-derived computer-aided detection (CAD) software for detecting and measuring lung nodules in CT scan images. [cited 2024 Jun 3]. Available online: https://www.nice.org.uk/guidance/dg55/chapter/1-Recommendations
20. MacMahon H, Naidich DP, Goo JM, et al. Guidelines for Management of Incidental Pulmonary Nodules Detected on CT Images: From the Fleischner Society 2017. Radiology 2017;284:228-43. [Crossref] [PubMed]
21. Baldwin DR, Gustafson J, Pickup L, et al. External validation of a convolutional neural network artificial intelligence tool to predict malignancy in pulmonary nodules. Thorax 2020;75:306-12. [Crossref] [PubMed]
22. Choi W, Oh JH, Riyahi S, et al. Radiomics analysis of pulmonary nodules in low-dose CT for early detection of lung cancer. Med Phys 2018;45:1537-49. [Crossref] [PubMed]
23. Ohno Y, Aoyagi K, Yaguchi A, et al. Differentiation of Benign from Malignant Pulmonary Nodules by Using a Convolutional Neural Network to Determine Volume Change at Chest CT. Radiology 2020;296:432-43. [Crossref] [PubMed]
24. Zhang Y, Jiang B, Zhang L, et al. Lung Nodule Detectability of Artificial Intelligence-assisted CT Image Reading in Lung Cancer Screening. Curr Med Imaging 2022;18:327-34. [Crossref] [PubMed]
25. McWilliams A, Tammemagi MC, Mayo JR, et al. Probability of cancer in pulmonary nodules detected on first screening CT. N Engl J Med 2013;369:910-9. [Crossref] [PubMed]
26. American College of Radiology. Lung CT Screening Reporting & Data System (Lung-RADS®). [cited 2024 Jun 3]. Available online: https://www.acr.org/Clinical-Resources/Reporting-and-Data-Systems/Lung-Rads
27. Callister ME, Baldwin DR, Akram AR, et al. British Thoracic Society guidelines for the investigation and management of pulmonary nodules. Thorax 2015;70:ii1-ii54. [Crossref] [PubMed]
28. Chung K, Jacobs C, Scholten ET, et al. Lung-RADS Category 4X: Does It Improve Prediction of Malignancy in Subsolid Nodules? Radiology 2017;284:264-71. [Crossref] [PubMed]
29. Ardila D, Kiraly AP, Bharadwaj S, et al. End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography. Nat Med 2019;25:954-61. [Crossref] [PubMed]
30. Schreuder A, Scholten ET, van Ginneken B, et al. Artificial intelligence for detection and characterization of pulmonary nodules in lung cancer CT screening: ready for practice? Transl Lung Cancer Res 2021;10:2378-88. [Crossref] [PubMed]
31. Jacobs C, van Rikxoort EM, Murphy K, et al. Computer-aided detection of pulmonary nodules: a comparative study using the public LIDC/IDRI database. Eur Radiol 2016;26:2139-47. [Crossref] [PubMed]
32. Silva M, Schaefer-Prokop CM, Jacobs C, et al. Detection of Subsolid Nodules in Lung Cancer Screening: Complementary Sensitivity of Visual Reading and Computer-Aided Diagnosis. Invest Radiol 2018;53:441-9. [Crossref] [PubMed]
33. Nair A, Screaton NJ, Holemans JA, et al. The impact of trained radiographers as concurrent readers on performance and reading time of experienced radiologists in the UK Lung Cancer Screening (UKLS) trial. Eur Radiol 2018;28:226-34. [Crossref] [PubMed]
34. Dellios N, Teichgraeber U, Chelaru R, et al. Computer-aided Detection Fidelity of Pulmonary Nodules in Chest Radiograph. J Clin Imaging Sci 2017;7:8. [Crossref] [PubMed]
35. Joy Mathew C, David AM, Joy Mathew CM. Artificial Intelligence and its future potential in lung cancer screening. EXCLI J 2020;19:1552-62. [Crossref] [PubMed]
36. O'Dowd E, Berovic M, Callister M, et al. Determining the impact of an artificial intelligence tool on the management of pulmonary nodules detected incidentally on CT (DOLCE) study protocol: a prospective, non-interventional multicentre UK study. BMJ Open 2024;14:e077747. [Crossref] [PubMed]
37. ClinicalTrials.gov. SCOOT: Sample Collection for DART. [cited 2024 Jun 3]. Available online: https://classic.clinicaltrials.gov/ct2/show/NCT05368298
38. Travis WD, Brambilla E, Riely GJ. New pathologic classification of lung cancer: relevance for clinical practice and clinical trials. J Clin Oncol 2013;31:992-1001. [Crossref] [PubMed]
39. Li Y, Chen D, Wu X, et al. A narrative review of artificial intelligence-assisted histopathologic diagnosis and decision-making for non-small cell lung cancer: achievements and limitations. J Thorac Dis 2021;13:7006-20. [Crossref] [PubMed]
40. Acs B, Rantalainen M, Hartman J. Artificial intelligence as the next step towards precision pathology. J Intern Med 2020;288:62-81. [Crossref] [PubMed]
41. Sakamoto T, Furukawa T, Lami K, et al. A narrative review of digital pathology and artificial intelligence: focusing on lung cancer. Transl Lung Cancer Res 2020;9:2255-76. [Crossref] [PubMed]
42. Ladbury C, Amini A, Govindarajan A, et al. Integration of artificial intelligence in lung cancer: Rise of the machine. Cell Rep Med 2023;4:100933. [Crossref] [PubMed]
43. Zhang YH, Jin M, Li J, et al. Identifying circulating miRNA biomarkers for early diagnosis and monitoring of lung cancer. Biochim Biophys Acta Mol Basis Dis 2020;1866:165847. [Crossref] [PubMed]
44. Pang S, Meng F, Wang X, et al. VGG16-T: A Novel Deep Convolutional Neural Network with Boosting to Identify Pathological Type of Lung Cancer in Early Stage by CT Images. Int J Comput Intell Syst 2020;13:771-80.
45. Moitra D, Mandal RK. Prediction of Non-small Cell Lung Cancer Histology by a Deep Ensemble of Convolutional and Bidirectional Recurrent Neural Network. J Digit Imaging 2020;33:895-902. [Crossref] [PubMed]
46. Guo Y, Song Q, Jiang M, et al. Histological Subtypes Classification of Lung Cancers on CT Images Using 3D Deep Learning and Radiomics. Acad Radiol 2021;28:e258-66. [Crossref] [PubMed]
47. Khosravi P, Kazemi E, Imielinski M, et al. Deep Convolutional Neural Networks Enable Discrimination of Heterogeneous Digital Pathology Images. EBioMedicine 2018;27:317-28. [Crossref] [PubMed]
48. Wang S, Wang T, Yang L, et al. ConvPath: A software tool for lung adenocarcinoma digital pathological image analysis aided by a convolutional neural network. EBioMedicine 2019;50:103-10. [Crossref] [PubMed]
49. Wei JW, Tafe LJ, Linnik YA, et al. Pathologist-level classification of histologic patterns on resected lung adenocarcinoma slides with deep neural networks. Sci Rep 2019;9:3358. [Crossref] [PubMed]
50. Khoo C, Rogers TM, Fellowes A, et al. Molecular methods for somatic mutation testing in lung adenocarcinoma: EGFR and beyond. Transl Lung Cancer Res 2015;4:126-41. [Crossref] [PubMed]
51. Girard N. Optimizing outcomes in EGFR mutation-positive NSCLC: which tyrosine kinase inhibitor and when? Future Oncol 2018;14:1117-32. [Crossref] [PubMed]
52. Shaw AT, Yeap BY, Mino-Kenudson M, et al. Clinical features and outcome of patients with non-small-cell lung cancer who harbor EML4-ALK. J Clin Oncol 2009;27:4247-53. [Crossref] [PubMed]
53. Lo Gullo R, Daimiel I, Morris EA, et al. Combining molecular and imaging metrics in cancer: radiogenomics. Insights Imaging 2020;11:1. [Crossref] [PubMed]
54. Coudray N, Ocampo PS, Sakellaropoulos T, et al. Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. Nat Med 2018;24:1559-67. [Crossref] [PubMed]
55. Zhang L, Chen B, Liu X, et al. Quantitative Biomarkers for Prediction of Epidermal Growth Factor Receptor Mutation in Non-Small Cell Lung Cancer. Transl Oncol 2018;11:94-101. [Crossref] [PubMed]
56. Liu Y, Kim J, Qu F, et al. CT Features Associated with Epidermal Growth Factor Receptor Mutation Status in Patients with Lung Adenocarcinoma. Radiology 2016;280:271-80. [Crossref] [PubMed]
57. Yamamoto S, Korn RL, Oklu R, et al. ALK molecular phenotype in non-small cell lung cancer: CT radiogenomic characterization. Radiology 2014;272:568-76. [Crossref] [PubMed]
58. Yoon HJ, Sohn I, Cho JH, et al. Decoding Tumor Phenotypes for ALK, ROS1, and RET Fusions in Lung Adenocarcinoma Using a Radiomics Approach. Medicine (Baltimore) 2015;94:e1753. [Crossref] [PubMed]
59. Jiang M, Sun D, Guo Y, et al. Assessing PD-L1 Expression Level by Radiomic Features From PET/CT in Nonsmall Cell Lung Cancer Patients: An Initial Result. Acad Radiol 2020;27:171-9. [Crossref] [PubMed]
60. Cicero M, Bilbily A, Colak E, et al. Training and Validating a Deep Convolutional Neural Network for Computer-Aided Detection and Classification of Abnormalities on Frontal Chest Radiographs. Invest Radiol 2017;52:281-7. [Crossref] [PubMed]
61. Domingues I, Pereira G, Martins P, et al. Using deep learning techniques in medical imaging: a systematic review of applications on CT and PET. Artif Intell Rev 2019;53:4093-160.
62. Halder A, Chatterjee S, Dey D. Adaptive morphology aided 2-pathway convolutional neural network for lung nodule classification. Biomed Signal Process Control 2022;72:103347.
63. Fiste O, Gkiozos I, Charpidou A, et al. Artificial Intelligence-Based Treatment Decisions: A New Era for NSCLC. Cancers (Basel) 2024;16:831. [Crossref] [PubMed]
64. Lococo F, Boldrini L, Diepriye CD, et al. Lung cancer multi-omics digital human avatars for integrating precision medicine into clinical practice: the LANTERN study. BMC Cancer 2023;23:1082. Erratum in: BMC Cancer 2023;23:1082. [Crossref] [PubMed]