Artificial intelligence enhances the management of esophageal squamous cell carcinoma in the precision oncology era

Abstract

Esophageal squamous cell carcinoma (ESCC) is the most common histological type of esophageal cancer with a poor prognosis. Early diagnosis and prognosis assessment are crucial for improving the survival rate of ESCC patients. With the advancement of artificial intelligence (AI) technology and the proliferation of medical digital information, AI has demonstrated promising sensitivity and accuracy in assisting precise detection, treatment decision-making, and prognosis assessment of ESCC. It has become a unique opportunity to enhance comprehensive clinical management of ESCC in the era of precision oncology. This review examines how AI is applied to the diagnosis, treatment, and prognosis assessment of ESCC in the era of precision oncology, and analyzes the challenges and potential opportunities that AI faces in clinical translation. Through insights into future prospects, it is hoped that this review will contribute to the real-world application of AI in future clinical settings, ultimately alleviating the disease burden caused by ESCC.

Key Words: Esophageal squamous cell carcinoma; Artificial intelligence; Deep learning; Machine learning; Precision tumor therapy

Core Tip: This article provides an overview of the current status of esophageal squamous cell carcinoma (ESCC) diagnosis and treatment, emphasizing the pivotal role of artificial intelligence (AI)-based predictive models in enhancing the precision of ESCC management. It outlines the existing challenges and opportunities associated with the integration of AI technologies in the diagnosis and treatment of ESCC. Furthermore, the article discusses the future research directions necessary to advance the practical application of AI in clinical settings, aiming to improve the accuracy and efficacy of ESCC care.

INTRODUCTION

Esophageal squamous cell carcinoma (ESCC) is the predominant histological type of esophageal cancer (EC), which accounts for 85% of all EC cases[1]. Despite rapid advancements in medical technology over the past decade, approximately 70% of patients cannot undergo surgery due to a high tumor burden in the advanced stages[2].

The improvement of endoscopic techniques provides excellent opportunities for early screening and cancer cure[3]. Nevertheless, the misdiagnosis rate in ESCC screening is approximately 40% in conventional white light endoscopy (WLI), posing obstacles to subsequent treatment[4]. Assessing tumor depth of infiltration and the presence of metastasis holds crucial guidance for selecting appropriate treatment modalities[5,6]. However, commonly used clinical methods in evaluating lymph node metastasis, such as endoscopic ultrasonography (EUS), positron emission tomography (PET), and computed tomography (CT) exhibit accuracies of only 66%, 68%, and 63%, respectively[7,8]. Furthermore, the individual variations in late-stage patients regarding responses to radiotherapy, chemotherapy, immunotherapy, and targeted therapy have become significant challenges in treatment[9,10]. Therefore, a method that comprehensively considers patients’ specific conditions, is urgently needed to achieve early detection, prognosis assessment, and precise prediction of treatment outcomes for ESCC.

With the progress of artificial intelligence (AI) algorithms and the growth of medical digital information, AI technology has been widely applied in various aspects of cancer research[11]. AI has demonstrated expert-level diagnostic and predictive capabilities in the precise detection and prognosis assessment of ESCC, showing significant potential for expanding and enhancing the efficiency of clinical work (Figure 1)[12]. As a crucial component of AI, machine learning (ML) has exhibited excellent results in extracting “semantic” information[13]. ML can extract the value from a large number of electronic health records, learning logical patterns in data and making predictions about disease development trends. Deep learning (DL), an emerging technology in the AI field, is currently primarily applied in the processing and analysis of imaging data[14]. Computer vision plays a vital role in this process, enabling computers to gain advanced understanding from visual inputs and providing information supporting medical imaging experts. Convolutional neural networks (CNN) are the most widely used tools in computer vision applications in the medical field[15]. Computer vision models can significantly enhance image gradation, identifying features the human brain cannot recognize. Conversely, the model’s input layer quantitatively processes image information, aiding in objective judgment and ensuring repeatability. Recently, DL architectures based on transformer have proven to be more efficient in natural language processing and computer vision tasks, holding the potential for efficient fusion of multimodal clinical information[16,17]. With the development of genomic sequencing technologies, the establishment of databases has also propelled the application of ML and DL techniques in identifying molecular features, such as gene mutations[18,19]. Through learning large-scale genomic data, the AI models can automatically identify gene expression patterns, mutation features, and gene interactions that are useful for disease diagnosis, enabling the inference of potential genetic variations and effectively facilitating the diagnosis and prediction of diseases.

Figure 1 Applications of artificial intelligence in esophageal squamous cell carcinoma. CT: Computed tomography; PET: Positron emission tomography.

This study used the keywords “artificial intelligence”, “deep learning”, “machine learning”, and “esophageal squamous cell carcinoma” to search relevant literature published up to April 2024 in databases such as PubMed, Embase, and Web of Science. Moreover, it provides a comprehensive overview of the applications of AI in the detection, diagnosis, prognosis assessment, and treatment of ESCC. Additionally, we analyzed the difficulties and challenges encountered in the clinical translation of AI and proposed potential solutions. We hope that through the review and analysis of previous research, we can contribute to advancing the exploration of AI in the field of cancer, particularly in ESCC.

APPLICATION OF AI IN SCREENING AND DIAGNOSIS OF ESCC

Cytological detection for screening of ESCC

In light of the lower invasiveness and higher cost-effectiveness of non-endoscopic cell sampling methods, innovative cytological devices such as expandable balloons and sponges have been developed for large-scale community screening[20]. However, cytological diagnosis relies on subjective judgment of physicians, and the shortage of cytology experts during large-scale screening significantly impacts the accuracy and reliability of diagnostic outcomes[21]. In 2021, Chinese scholars introduced a novel esophageal cell collector utilizing sponge-like polymer materials for esophageal specimen collection, integrating AI technology to detect of abnormal cells on glass slides[22]. Results from studies conducted across multiple centers indicate that AI-assisted sponge cytology examinations exhibit high sensitivity and specificity in screening for abnormal proliferation of esophageal squamous epithelium[22,23]. Thus, Gao et al[24] integrating cytological and epidemiological features, utilized a CNN model to identify the risk of developing ESCC. The validated model achieved an area under the curve (AUC) value of 0.96, which was comparable to the diagnostic proficiency of cytologists. Importantly, the application of an AI-based fully automated screening method can additionally avert 92.8% of unnecessary endoscopic examinations, marking a milestone in EC screening[25].

Tumor biomarkers for screening of ESCC

Tumor biomarkers (TMs) are substances released by tumor cells or induced by host cells during tumor initiation and development[26]. Studies have shown a significant potential of blood and exhaled breath samples in identifying high-risk populations for ESCC[27]. However, the sensitivity and specificity of these markers, such as volatile organic compounds, circulating miRNA, and circular RNA, that can be used for ESCC screening are not ideal[28-31], and clinically applicable ESCC-TMs remain limited.

Utilizing AI methods to mine large-scale omics data can efficiently identify TMs with high specificity. Furthermore, constructing tumor classifier models using AI technology based on known or unknown biomarkers can prospectively assess the diagnostic value of these features in ESCC, enabling the possibility of an ESCC screening strategy based on TMs. Yuan et al[32] combined ML algorithms with serum lipidomics to identify a set of lipid TMs for screening ESCC, with a sensitivity of 90.7% in an independent validation cohort. Zhu et al[33] utilized a random forest algorithm for metabolomic analysis of plasma exosomes, using the selected four potential markers to construct an ML model. The results showed an AUC of 0.98 in predicting recurrence risk in patients with ESCC. Wang et al[34] developed a novel method for encoding untargeted metabolomic data obtained through liquid chromatography-mass spectrometry, the MetImage. Using multichannel images encoded by MetImage, a CNN-based DL model that accurately identified ESCC patients, exhibiting sensitivities, specificities, and AUC values of 85%, 92%, and 0.95, respectively.

Endoscopic detection methods for ESCC

The primary method for screening ESCC and detecting precancerous lesions is endoscopic examination[3]. Several studies have shown that regular endoscopic screening for high-risk patients with ESCC, with the option for endoscopic or surgical intervention as needed, can greatly improve prognosis[35,36]. Nonetheless, inexperienced physicians may overlook subtle changes in lesions during the examination, leading to missed diagnoses. Due to workload or fatigue, even experienced doctors may misdiagnose or overlook diseases.

DL models automatically extract complex features from endoscopic images through a multilayer network of structures. These models can detect subtle lesions or abnormal patterns and model intricate relationships through nonlinear transformations. This efficient feature learning and data processing ability allows AI models to excel in the diagnosis, differentiation assessment, and prediction of invasion depth of ESCC under endoscopy, significantly enhancing the diagnostic efficiency of clinicians (Table 1). Additionally, Yuan et al[37] developed an AI system based on deep CNN (DCNN) to identify various ESCC imaging patterns during endoscopy. The AI system can make a judgment in 17 microseconds, reducing diagnostic time by 300 times. Regarding diagnostic performance, the AI system’s judgments in all endoscopic imaging patterns were satisfactory, with sensitivity and specificity ranging from 92.5% to 99.7% and 78.5% to 89.0%, respectively. Zhao et al[38] reported that the DL model showed an accuracy approximately 15% higher than that of junior physicians in lesion classification. Li et al[39] integrated a fully automatic AI system into narrow-band imaging (NBI) and WLI. Following AI-guided results, the endoscopists’ specificity and accuracy in diagnosing ESCC lesions improved to 96.7% and 94.9%, respectively. These research findings indicate that AI serves as a diagnostic auxiliary tool and potentially functions as an effective training system, providing expert-level guidance for junior physicians to make accurate judgments.

Table 1 Artificial intelligence in endoscopic examination of esophageal squamous cell carcinoma.

Ref.	Data sets	Endoscopic modality	Main study aim	AI algorithms	Best result
Yuan et al[37], 2022	Image data set: 53933; images from 2621 patients; video data set: 142; videos from 19 patients	WLI, NM-NBI, ME -NBI	Detection of superficial ESCC	DCNN	Image data set: Sensitivity: 92.5%-99.7%; specificity: 78.5%-89.0%; AUC: 0.906-0.989; video data set: Sensitivity: 89.5%-100%; specificity: 73.7%-89.5%
Zhao et al[38], 2019	Image data set: 1383; images from 219 patients	ME-NBI	Classification of IPCLs	FCN	Mean accuracy of senior IPCL: 92.0%; mid-level IPCL: 82.0%; junior IPCL: 73.3%
Li et al[39], 2021	Image data set: 5367 images	NM-NBI	Detection of early ESCC	FCN	AUC: 0.9761; sensitivity: 91.0%; specificity: 96.7%; accuracy: 94.3%
Fukuda et al[41], 2020	Image data set: 28333 images	ME-NBI	Detection and characterization of early ESCC	CNN	Detection: Sensitivity: 91.0%; specificity: 51.0%; accuracy: 63.0%; characterization: Sensitivity: 86.0% specificity: 89.0%; accuracy: 88.0%
Liu et al[42], 2022	Image data set: 13083; WLI images from 1239 patients	WLI	Detection and delineation of ESCC margins	DCNN	Detection: Accuracy: 85.7% (internal validation) and 84.5% (external validation); delineation: Accuracy: 93.4% (internal validation) and 95.7% (external validation)
Ikenoyama et al[44], 2021	Image data set: 7301; images from 667 patients	LCE	Predict multiple Lugol-voiding lesions	GoogLeNet deep neural network	Sensitivity: 84.4%; specificity: 70.0%; accuracy: 76.4%
Yuan et al[47], 2023	Image data set: 10047; images from 1112 patients; video data set: 140; videos from 1183 lesions	ME-NBI	Detection and delineation of ESCC margins	DCNN	Detection: Accuracy: 92.4% (internal validation) and 89.9% (external validation); delineation: Accuracy: 88.9% (internal validation) and 87.0% (external validation)
Shimamoto et al[48], 2020	Image data set: 23977; images from 909 patients; video data set: 102 videos	NM-NBI	Assessment of tumor infiltration depth	CNN	Sensitivity: 50.0%; specificity: 99.0%; accuracy: 87.0%
Wang et al[49], 2023	ME-BLI data set: 2887; images from 246 patients; ME-NBI data set: 493; images from 81 patients	ME-BLI, ME-NBI	Identification of IPCL	R-CNN	Recall: 79.25%; precision: 75.54%; F1-score: 0.764; mean average precision: 74.95%
Zhang et al[53], 2023	Image data set: 5119; images from 581 patients; video data set: 33 videos	ME-NBI	Infiltration depth prediction	AI-IDPS	For differentiating SM2-3; lesions: Image data set: Sensitivity: 85.7%; specificity: 86.3%; accuracy: 86.2%; video data set: Sensitivity: 87.5%; specificity: 84.0%; accuracy: 84.9%
Yuan et al[54], 2022	Image data set: 7094; images from 685 patients	ME-NBI	Classification of IPCLs	DCNN	Accuracy (internal validation): 91.3%; accuracy (external validation): 89.8%

CNN: Convolutional neural networks; DCNN: Deep convolutional neural networks; FCN: Fully convolutional network; R-CNN: Region convolutional neural network; IPCLs: Intrapapillary capillary loops; WLI: White light image; LCE: Lugol chromoendoscopy; NM-NBI: Non-magnified endoscopy with narrow-band imaging; ME-NBI: Magnifying endoscopy with narrow-band imaging; ME-BLI: Blue laser imaging; AI: Artificial intelligence.

WLI: Common methods for EC screening are observation of suspicious lesions and collection of biopsy samples under WLI. However, early EC lesions may manifest as subtle mucosal red changes that can be easily overlooked during WLI examinations[4]. However, AI algorithms are adept at recognizing subtle changes in endoscopic images. Utilizing techniques, such as data augmentation, multiscale and multiview analyses, and transfer learning, these algorithms conduct in-depth image analysis and make judgments that are more sensitive than those of endoscopy experts (sensitivity: 90.8% compared with 82.5%)[37]. To further reduce the risk of oversight, AI algorithms can also analyze video streams directly, automatically detecting and annotating potential lesion areas and providing real-time feedback. In a randomized controlled trial by Yuan et al[40], the diagnostic performance of the AI system was validated in real clinical settings across 12 centers. The researchers found that applying a video-based lesion detection AI system reduced the overall lesion miss rate and miss rate per patient by 5.0% and 3.2%, respectively[41].

The accurate delineation of the margins of early ESCC is crucial in evaluating endoscopic surgical margins. Liu et al[42] constructed a DCNN model that can outline lesion edges under WLI with an accuracy of 98%, surpassing the 95% accuracy of endoscopy experts. By combining endoscopic and pathological features, the AI model can predict the risk of lymph node metastasis, providing a valuable reference for tumor staging and subsequent treatment plan selection.

Lugol chromoendoscopy: Lugol chromoendoscopy (LCE) is performed based on the principle that abnormal esophageal mucosal cells, lacking glycogen, do not take up iodine stain. By assessing the staining pattern of the esophageal mucosa and performing targeted biopsies, LCE can further reduce the incidence of EC by approximately 70%[43]. When LCE videos are input into the DL model by Yuan et al[37], the diagnostic accuracy, sensitivity, and specificity for ESCC can reach 94.7%, 100%, and 89.5%, respectively. However, iodine staining may induce allergies and exacerbate hyperthyroidism, which limits the widespread application of LCE. Therefore, Ikenoyama et al[44] developed a diagnostic system based on the GoogLeNet deep neural network, that can predict Lugol-voiding lesions in the esophagus from non-iodine-stained endoscopic images and identify high-risk patients with ESCC.

NBI endoscopy: Although it may not be considered the optimal technique, NBI has emerged as a new endoscopic technology for EC screening. Meta-analysis shows that NBI has a sensitivity of 88% in identifying tumor patients, with the remaining 12% missed detection rate[45]. Inexperienced endoscopists who used NBI technology to identify ESCC may have a sensitivity as low as 53%[46]. Yuan et al[47] integrated an AI system into endoscopic devices for real-time detection of ESCC and precancerous lesions, as well as delineating tumor boundaries. The system demonstrated satisfactory performance, with lesion detection and boundary delineation accuracies of 91.4% and 85.9%, respectively, which was comparable to experienced endoscopists.

A crucial indication for esophagectomy is tumors with a risk of SM2-3 invasion > 25%. Shimamoto et al[48] developed a CNN-based tumor infiltration depth assessment system, which can analyze continuous image data from NBI videos in real-time, identify tumor lesions, and differentiate between SM1 and SM2-3 tumors. The system’s accuracy, sensitivity, and specificity were 87.3%, 50.0%, and 98.7%, respectively, surpassing the diagnostic proficiency of medical experts (84.7%, 44.9%, and 96.9%, respectively).

Magnifying endoscopy with NBI: The installation of a high-magnification lens on an endoscope, combined with NBI technology, allows a clearer observation of the fine structures of the esophageal mucosa epithelium and the morphological features of intrapapillary capillary loops (IPCLs), facilitating the assessment of tumor differentiation and infiltration depth. Wang et al[49] designed the PSA-HRNetV2p model based on region CNN, enabling automatic localization and identification of IPCL in magnifying endoscopy with NBI (ME-NBI). The Japan Esophageal Society proposed the JES classification based on different IPCL morphologies, categorizing IPCL into types A (normal or inflammatory tissue) and B (tumor tissue with varying degrees of infiltration)[50]. Zhao et al[38] introduced a dual-label fully convolutional network model that can identify IPCL types, with an accuracy of 87% in lesion classification.

Moreover, ME-NBI is an effective method for predicting the infiltration depth of ESCC, with an accuracy of 76% in distinguishing between tumor mucosal/submucosal infiltration[51]. The accuracy of assessing the invasion depth of ESCC can increase to 79% when used in conjunction with EUS[52]. Zhang et al[53] incorporated potential endoscopic visual features and developed the AI-based ESCC infiltration depth assessment system AI-IDPS. The accuracy of endoscopists in assessing the infiltration depth of ESCC can be further increased by 5% with the assistance of AI-IDPS. When evaluating tumor infiltration depth using the JES classification, the DCNN model by Yuan et al[54] achieved a comprehensive accuracy of approximately 90% in diagnosing IPCL subtypes and infiltration depth.

Tissue pathology-assisted diagnosis and staging of ESCC

The gold standard for diagnosing ESCC is histopathological examination[3]. However, pathological diagnosis heavily relies on the individual experience of pathologists, and variations among different observers may lead to different diagnostic results or even misdiagnosis[55]. By using AI for the in-depth analysis of complex pathological data enables detailed quantitative analysis of pathological slide images, such as cell counting and morphological feature extraction. Subsequently, this approach provides an objective basis for tumor diagnosis, staging, and prognostic evaluation[56-58]. The CNN model by Yang et al[56] quantifies cell nucleus morphology in pathological images, accurately classifying the five stages of ESCC tissue transformation, with an accuracy of up to 99%. Kouzu et al[57] used a DL classifier to classify the proliferative response of connective tissue in pathological images, which achieved precise individualized prognosis prediction.

With the development of digital pathology, the advent of whole slide imaging (WSI) further promotes the application of AI in pathology[55]. In histological-based diagnosis, every cell in a sample can potentially be a tumor cell. Relying on expert identification and annotation of pixel regions that contain tumor cells for building large-scale datasets for model training is time-consuming and impractical, leading to the application of self-supervised methods[59]. Self-supervised learning models can autonomously learn from WSI information, which concludes by extracting potential features and patterns in the WSI without relying on pathologists’ annotations[58,60]. This fully automated architecture is suitable for large-scale image analysis, providing objective and reproducible results, and holding great potential in tumor pathology.

Image data-assisted diagnosis and staging of ESCC

AI has been widely applied in medical imaging, demonstrating significant advantages in improving image quality and providing automated diagnoses. Using AI for CT image reconstruction helps identify and suppress noise from acquired projection signals, optimizing image clarity and contrast, which allows for more accurate observation of the morphological features of tumors[61]. Conversely, AI algorithms enable the quantitative analysis of medical imaging, providing objective diagnostic information[62]. CT scans are a routine radiological method to assess lymph node status preoperatively in patients with ESCC[8]. However, in clinical practice, > 20% of patients with ESCC may have occult lymph node metastasis, where the size and morphology of these lymph nodes do not show obvious changes in CT imaging. Neglecting these in preoperative staging examinations can directly lead to rapid tumor recurrence immediately postoperatively[63,64]. Zhang et al[65] found that ML models have significant advantages in detecting smaller occult lymph node metastases. This AI-based non-invasive predictive method effectively reduces the risk of postoperative recurrence and avoids unnecessary lymph node dissection in low-risk patients. Xie et al[66] input CT image radiomics features into a support vector machine (SVM) model to construct a single lymph node metastasis classification model, with an accuracy of 94.7%. These AI tools aid clinicians in making decisions regarding surgical approaches and the extent of lymph node dissection.

APPLICATION OF AI IN THE TREATMENT OF ESCC

Surgical treatment

Compared with traditional surgical approaches, endoscopic resection and ablation techniques have advantages, such as minimal postoperative trauma, rapid recovery, and fewer complications, making them the first-line treatment for early-stage ESCC[67]. However, when the tumor infiltrates more than one-third of the submucosal layer with a depth > 200 μm, the risk of lymph node metastasis can be as high as 50%, necessitating surgical intervention[68]. As mentioned earlier, computer vision can identify tiny details and patterns in images from endoscopy, histopathology, and radiology, which can effectively assess tumor infiltration depth, extent of involvement, and lymph node metastasis. Accurate preoperative staging aids in selecting appropriate surgical approaches. However, clinical practice indicates that patients at the same tumor node metastasis (TNM) stage may have different risks of early postoperative recurrence, which may be attributed to the heterogeneity of ESCC, influencing biological behaviors, such as tumor cell migration and invasion[69]. Yang et al[70] integrated TNM staging, clinicopathological parameters, and tumor molecular markers to construct an SVM-based nomogram model, which predicts the risk of postoperative recurrence in patients and facilitates personalized and precise treatment strategies.

Robot-assisted esophagectomy

In 2003, robot-assisted minimally invasive esophagectomy (RAMIE) was initially introduced as a minimally invasive approach for the treatment of EC using robotic surgical systems. The use of high-definition 3D imaging technology provides an enlarged surgical view, which enhances the precision of the procedure. Additionally, the flexibility of robotic arms and tremor filtration technology further improved the stability and accuracy of the surgery. Several studies have validated its feasibility and safety. Van der Sluis et al[71] conducted a randomized controlled trial that compared RAMIE with open thoracoscopic esophagectomy (OTE) in terms of postoperative complications and other factors. They found that RAMIE and OTE performed equally well in achieving radical tumor resection, meeting the highest standards of contemporary practice. Compared with OTE, patients treated with RAMIE had lower rates of surgery-related complications and cardiopulmonary complications, along with better postoperative recovery. Yang et al[72] conducted a prospective comparative study of the postoperative outcomes of patients with ESCC treated with RAMIE and conventional minimally invasive esophagectomy. They demonstrated the safety of RAMIE in ESCC tumor treatment and also highlighted the significant advantages of RAMIE in identifying and dissecting metastatic lymph nodes. However, the effectiveness of RAMIE is highly influenced by the learning curve, and the postoperative outcomes also depend on the surgeon’s experience and expertise accumulation[73].

Neoadjuvant therapy

The standard treatment approaches for locally advanced resectable ESCC are neoadjuvant chemotherapy (NAC) and neoadjuvant chemoradiotherapy (NACRT)[7,67,74]. However, predicting patient response to NAC and NACRT is challenging due to tumor heterogeneity. Demographic characteristics, medical history, clinical staging, and other factors may contribute to treatment-related side effects and tumor recurrence after treatment[10]. Thus, establishing a non-invasive and accurate method for predicting patient responses to treatment before initiating therapy is crucial for guiding treatment decisions.

Kawahara et al[75] reported that DL models can predict pathological complete response (pCR) after NACRT by extracting and analyzing treatment-related features from endoscopic images. Hu et al[76] used CT radiomic features for training ML models, with an accuracy of 77.1%. Combining CT radiomic with clinical features, Wang et al[77] used ML algorithms to predict the pCR of primary tumor lesions and lymph node metastases. Subsequently, the combination of features from different dimensions significantly improved the predictive performance of the model. Interpreting the impact of genetic variations is the key to understanding an individual’s sensitivity to treatment regimens[78]. Sasagawa et al[79] developed an RF prediction model using immunogenomics and clinical information, which provided individualized predictions of patient responses to NAC and identified molecular markers closely associated with NAC response through feature analysis, opening new avenues for early assessment of patient treatment sensitivity.

Treatment for unresectable ESCC

Concurrent chemoradiotherapy: For patients with unresectable ESCC or ESCC who cannot tolerate surgery, the 2017 Japanese EC Practice Guidelines recommend curative concurrent chemoradiotherapy[67]. Huang et al[80] integrated pre-treatment CT imaging and EHR using ML methods to predict the objective response rates and progression-free survival (PFS) of patients with unresectable ESCC undergoing different chemoradiotherapy regimens. Moreover, this helps in formulating personalized treatment strategies. Li et al[81] designed a 3D-DLRM model based on ResNet34, which, by exploring high-dimensional imaging features in CT images, predicts the response of patients receiving radiation therapy to different radiotherapy plans, radiation doses, and radiation fields. Compared with less commonly used biomarkers and expensive PET-CT scans, CT-based DL methods contribute to achieving more accurate and convenient personalized treatment[82,83]. Clinical and radiomic features are high-risk factors associated with the prognosis of patients with ESCC. Cui et al[84] extracted clinical variables and contrast-enhanced CT images from patients, using a combined model to assess PFS and overall survival in patients with advanced ESCC receiving chemoradiotherapy. This achieved better performance than single-modal models.

Targeted therapy and immunotherapy: Due to individual differences in traditional chemoradiotherapy regimens, the treatment most effective in each patient is difficult to provide. Intratumoral genetic heterogeneity and clonal evolution also contribute to the development of chemotherapy resistance[69]. The local recurrence rate after treatment remains at 40%-60%, with a < 30% 5-year survival rate[64]. Moreover, traditional chemoradiotherapy lacks specificity and may damage not only tumor tissues but also normal tissue cells during treatment[85]. The development of targeted therapy and immunotherapy has brought survival benefits to patients[86]. However, only a few drugs are available for patients with ESCC patients[87-90]. New treatment targets, new targeted drugs, and screen biomarkers for immunotherapy should be identified to improve the prognosis of patients with ESCC[86]. Intratumoral heterogeneity (ITH) is a potential target for the development of combination treatment strategies[91]. Moreover, DL technology is recommended for integrating and analyzing multi-omics data, providing a powerful tool for dissecting ITH[69]. In addition, considering the intertumoral heterogeneity of tumors, building personalized molecular targeted drugs for specific patients also helps optimize treatment strategies[92]. Chen et al[93] combined metabolomics with ML algorithms to identify three potential therapeutic targets, SLC1A5, SLC7A5, and SLC16A10, in ESCC samples. Overall, AI models are expected to provide more precise and efficient tools for tumor immunotherapy and targeted therapy, which will contribute to the advancement of precision medicine in cancer.

APPLICATION OF AI IN SURVIVAL PREDICTION FOR ESCC

Personalized prognostic evaluations contribute to understanding patient survival and treatment outcomes, enabling more informed treatment decisions. By analyzing extensive data, including genomics, pathology, and laboratory examinations, AI can identify key factors associated with patient prognosis, offering novel insights and directions for disease management. The tertiary lymphoid structure (TLS) is an ectopic lymphoid organ formed by the aggregation of various immune cells[94]. Ling et al[95] demonstrated that patients with ESCC with TLS-mature structures exhibit a 5-year disease-free survival rate approximately 20% higher than TLS-negative patients. A classification model based on inception-resnet-v2 has an accuracy of 95.3% in evaluating TLS maturity. Characterizing genomic information is crucial for understanding the functionality of each genomic component in the organism, which plays an important role in individualized clinical outcome assessments[96]. Li et al[97] utilized five ML algorithms and identified critical molecules that intersect in the prediction of ESCC prognosis, highlighting Stratifin protein as a key prognostic marker. Additionally, log-rank test results demonstrated the accuracy of Stratifin-based risk stratification in accurately characterizing the median survival time of patients.

Building AI predictive models based on potential factors can quantify patient prognosis, facilitating more precise and personalized treatment strategies. Table 2 highlights that AI models are currently extensively used for feature extraction and analysis in clinical indicators, CT images, and histopathological images, achieving outstanding results in predicting the survival of patients with ESCC[98-103].

Table 2 Application of artificial intelligence in survival prediction for esophageal squamous cell carcinoma.

Ref.	Sample size	Data sources	AI algorithms	Best result	Highlight
Zhang et al[98], 2023	Training: 1954 cases; testing: 487 cases	Clinicopathological and laboratory information	Rpart, Elastic Net, RF, GBM, GLMboost, and CoxPH	AUC (1-year overall survival): 0.725; AUC (3-year overall survival): 0.0.720; AUC (5-year overall survival): 0.752	Prognosis evaluation after ESCC resection. First study comparing different AI algorithms’ performance in predicting the survival prognosis of ESCC
Zhang et al[99], 2023	Training: 3612 cases; validation: 1204 cases testing: 1204 cases	Demographic, clinicopathological, and treatment information	DeepSurv	C-index: 0.732; AUC (3-year overall survival): 0.805; AUC (5-year overall survival): 0.825	Survival prediction of ESCC. Creating a web-based tool for individual survival outcomes prediction
Ling et al[100], 2023	Training: 4062 cases; testing: 1015 cases	Demographic, clinicopathological, and treatment information	WOA-XGBoost	AUC: 0.918; accuracy: 88.1%	Prognosis evaluation of stage IB-IVA ESCC after clinical treatment. The integration of IDPC and WOA-XGBoost realizes a more accurate prediction
Wang et al[101], 2021	Training: 122 cases; testing: 31 cases	CT and histopathological image	Regularized Cox, Weibull Cox, RSF, and DeepSurv	C-index: 0.694	Prognosis evaluation after ESCC resection. Successful survival analysis through combining multiple modal images together
Liu et al[102], 2020	Training: 79 cases; testing: 71 cases	Transcriptome profile	RF, SVM	Specificity: 98.1%; sensitivity: 79.0%	Prognosis evaluation of patients with ESCC. Identifying novel genetic markers reflecting the prognosis of ESCC
Yi et al[103], 2020	Training: 182 cases; testing: 179 cases	Tumor stemness index provided by one-class logistic regression	Combination of Lasso and Cox regression	AUC (1-year survival time): 0.947; AUC (3-year survival time): 0.760; AUC (5-year survival time): 0.770	Survival prediction of ESCC. First study predicting the prognosis using the tumor intrinsic stemness pattern

RF: Random forest; GBM: Gradient boosting machine; RSF: Random survival forest; SVM: Support vector machine; ESCC: Esophageal squamous cell carcinoma; CT: Computed tomography; AI: Artificial intelligence; AUC: Area under the curve.

CHALLENGES AND OPPORTUNITIES

AI models, with their exceptional feature extraction and learning capabilities, have demonstrated high sensitivity and specificity in the diagnosis and treatment of ESCC. However, successfully integrating AI into clinical applications for ESCC remains challenging (Figure 2A).

Figure 2 Challenges and opportunities of artificial intelligence in the management of esophageal squamous cell carcinoma. A: Examples of challenges and opportunities encountered by artificial intelligence (AI) in the field of esophageal squamous cell carcinoma; B: The fundamental architecture of federated learning (FL) in medical applications. FL integrates data from multiple healthcare institutions, enabling collaborative modeling. Initially, a central server establishes a base model and communicates the model's structure and parameters to each healthcare institution. Institutions perform local training using their respective data and return the results to the central server. Through the aggregation of data flows from various institutions, a high-quality shared global model is constructed; C: Brief overview of the integration of multimodal data. Multimodal AI consolidates and analyzes audio, image, and text information to build high-precision models. AI: Artificial intelligence.

We must consider issues, such as data bias caused by implicit or explicit biases existing in healthcare systems. Biases related to specific patient groups, sex, and ethnicity can influence the training and deployment of AI models[104]. Particularly in tumor prediction, AI models constructed based on minority populations lead to insufficient predictive performance and difficulties in generalization. Therefore, digitalization of the healthcare sector should be further enhanced, collaboration and communication should be facilitated among different medical institutions, and the application of healthcare AI should be expanded through large-scale data training and model optimization. In 2016, Google proposed the “federated learning” approach, enabling the aggregation of data from distributed health data clients, providing a feasible way to connect electronic medical record data across healthcare institutions (Figure 2B)[105]. Recently, Huang et al[106] curated a large dataset called OpenPath that used pathology images shared by clinical professionals on public forums, such as Twitter. This dataset was designed for training multimodal models with image and text understanding capabilities. They proposed that the development of social media platforms, such as Twitter, significantly facilitates the open sharing of medical information. Embracing this valuable yet underutilized opportunity is crucial for fostering the continued development of medical AI.

Clinical decisions are intricate. Relying on either single-image features or multimodal features alone may not adequately reflect the patient’s prognosis. A more comprehensive and accurate patient profile can be obtained by merging data from different modalities, enabling the formulation of more precise treatment plans[107]. However, current predictive models for ESCC often rely on single-modal data, overlooking crucial interactive information between modalities. In 2021, OpenAI initiated AI multimodal learning, which aimed to handle and correlate information from various modalities (Figure 2C)[108]. Integrating histopathological, radiological images, and even genomic information into models can provide supplementary evidence for tumor diagnosis and prediction, resulting in performance superior to unimodal models[109-111]. However, selecting effective fusion strategies to address challenges, such as the varying strengths of different modalities and noise due to misalignment, remains a major challenge. Thus, this will be a key focus of future AI research in the medical field.

The inherent “black box” nature of AI algorithms poses challenges in terms of interpretability, which limits the applicability of models[112]. In medicine, the reliability of model decision processes and the accuracy of prediction outcomes are equally crucial. Thus, improving decision transparency increases patient trust and enhances healthcare quality. In 2015, the United States Defense Advanced Research Projects Agency introduced the Explainable AI (XAI) program[113]. In 2016, the European Union enacted the General Data Protection Regulation (GDPR), incorporating provisions for the “right to explanation” concerning automated decisions by AI algorithms[114]. Addressing ethical concerns in the design and implementation of AI models, XAI provided interpretable insights into the decision-making process[115]. Lee et al[116] developed an XAI algorithm based on class activation mapping for localizing acute intracranial hemorrhage. Ming et al[117] introduced a visualization method named RuleMatrix based on rule extraction, aiding users in understanding how ML models classify tumor cells. By offering accessible explanations of AI system inferences, XAI contributes to building fair, transparent, and accountable AI decision systems.

The privacy and data security concerns associated with AI model training and validation must also be considered. Insufficient data encryption measures can expose stored and transmitted data to unauthorized access risks. Additionally, a lack of effective access control can lead to information leakage and data misuse[118]. Therefore, a range of supplementary measures should be implemented to ensure effective data utilization while protecting individual privacy. To ensure the legality and security of patient health data, healthcare institutions must comply with data protection regulations, such as the GDPR[119] and the Health Insurance Portability and Accountability Act[120].

Furthermore, most current research evaluates the effectiveness of models in retrospective datasets, overlooking the characteristic variability of the clinical environment, thereby limiting their clinical application. Thus, subsequent efforts should involve extensive testing of AI models in clinical settings to better comprehend and interpret the fluctuations and changes in data. Utilizing AI-assisted diagnostic tools for prospective assessments in real clinical environments will more convincingly demonstrate the model’s genuine clinical effectiveness in the face of dynamic conditions.

CONCLUSION

The era of AI has arrived, and the application of AI technologies is ushering in unprecedented transformations and opportunities in medicine. The potential of AI should be further explored in the diagnosis and personalized treatment of ESCC. While ensuring the safety and reliability of the technology, the utilization of AI should also be enhanced, assisting clinicians in decision-making, thereby achieving more efficient clinical diagnosis and treatment.