Artificial intelligence applications in exposed endoscopic full-thickness resection vs thoracoscopic surgery for complex esophageal subepithelial lesions: Narrative review

Abstract

Complex esophageal subepithelial lesions (SELs; ≥ 35 mm or cervical location) originating from the muscularis propria challenge standard submucosal tunneling endoscopic resection due to risks of tunnel collapse, perforation, and retrieval failure. Exposed endoscopic full-thickness resection (EFTR) and thoracoscopic surgery (TS) offer viable alternatives, yet their complexity supports artificial intelligence (AI) integration for lesion characterization, workflow recognition, anatomical guidance, and complication prediction. This narrative review synthesizes evidence to February 2026 using a three-tier framework: Direct (esophageal SEL-specific EFTR/TS), adjacent (esophageal/upper gastrointestinal endoscopy), and transferable (surgical/robotic AI). AI-enhanced endoscopic ultrasound improves SEL accuracy (89%-95%); real-time phase recognition achieves 92%; robotic systems aid TS planning. SEL-specific prospective trials remain essential.

Key Words: Artificial intelligence; Endoscopic full-thickness resection; Thoracoscopic surgery; Subepithelial lesions; Esophageal tumors; Machine learning; Endoscopic ultrasound; Computer-aided detection

Core Tip: Artificial intelligence holds promise for esophageal subepithelial lesion management through endoscopic ultrasound characterization, exposed endoscopic full-thickness resection workflow support, and thoracoscopic surgery navigation; evidence is primarily adjacent/transferable, warranting procedure-specific validation.

INTRODUCTION

Esophageal subepithelial lesions (SELs) arising from the muscularis propria-such as leiomyomas and gastrointestinal stromal tumors-are increasingly detected incidentally via endoscopy or endoscopic ultrasound (EUS)[1-5]. While small asymptomatic lesions warrant observation, complex cases (≥ 35 mm, cervical/proximal, irregular borders, or symptomatic) demand intervention. Submucosal tunneling endoscopic resection succeeds for smaller muscularis propria tumors but falters in complex anatomy due to tunnel collapse, limited space, scope instability, and piecemeal retrieval[4,6-8].

Exposed endoscopic full-thickness resection (EFTR) provides a tunnel-free endoscopic option, whereas thoracoscopic surgery (TS) excels for posterior/extraluminal lesions[9-14]. Both approaches are effective, but each carries technical risks that may benefit from artificial intelligence (AI) support. Both yield high success but carry technical risks amenable to AI via convolutional neural networks, U-Net segmentation, and vision transformers for imaging and workflow analysis[15-19]. This review evaluates AI across the EFTR-TS spectrum, distinguishing direct from indirect evidence.

SEARCH STRATEGY

Literature search to February 28, 2026, spanned PubMed/MEDLINE, EMBASE, Scopus, Web of Science, Google Scholar, and gray sources. Terms combined AI (machine learning, deep learning, computer-aided detection), procedures (EFTR, TS), anatomy (esophagus, SEL), and pathologies (leiomyoma, gastrointestinal stromal tumor). Included: English clinical trials, cohorts, validations, reviews/meta-analyses (2017-2026). Excluded: Case reports, non-peer-reviewed abstracts. Yield: 1478 hits; 236 full-texts screened; 39 included (multicenter AI n = 14, EUS/CADe n = 9, phase recognition n = 10, robotics/thoracoscopy n = 6).

CLINICAL BACKGROUND: EFTR VS TS

EFTR involves saline lift, marking, incision, dissection, transection, retrieval, and closure (over-the-scope clip)[9,10]. Granata et al[10] report 100% success in upper gastrointestinal SELs, though strictures reach 20%-50% cervically. TS uses 3-4 ports for enucleation, ideal posteriorly[11-14]. In Zhengzhou RCT (n = 92; EFTR n = 46, TS n = 46), EFTR achieved R0 86.7% (vs TS 75.6%, P = 0.21), shorter nasogastric duration (5.6 vs 10.7 days, P < 0.001), and hospital stay (6.8 days vs 12.3 days, P < 0.001). EFTR risks perforation/leak (5%); TS risks chylothorax/fistula (6.5%-8.9%). Table 1 summarizes non-AI outcomes, framing AI needs in precision and safety[20].

Table 1 Clinical outcomes: Exposed endoscopic full-thickness resection vs thoracoscopic surgery, n (%)/mean ± SD.

Outcome	Exposed EFTR (n = 46)	TS (n = 46)	P value
Technical success	46 (100)	45 (97.8)	0.41
En bloc/R0 resection (%)	86.7	75.6	0.21
NG tube (days)	5.6 ± 2.1	10.7 ± 4.3	< 0.001
Hospital stay (days)	6.8 ± 1.9	12.3 ± 3.7	< 0.001
Recurrence (28 months) (%)	0	2.2	0.31

EFTR: Exposed endoscopic full-thickness resection; TS: Thoracoscopic surgery.

LESION DETECTION AND CHARACTERIZATION

Preoperative EUS guides SEL assessment, but operator variability persists[2-4]. Direct SEL studies show AI-EUS sensitivity 89%-95%, accuracy 92%[2,4,21]. Adjacent upper-GI evidence supports texture/contour analysis[4,15,22]. AI refines EFTR/TS selection for equivocal lesions[2,22].

WORKFLOW AND PHASE RECOGNITION

Workflow and phase recognition research offers robust adjacent evidence for AI procedural support in EFTR. Multicenter esophageal endoscopic submucosal dissection studies demonstrate AI phase identification accuracy exceeding 89%, with improved trainee performance via AI-annotated videos[12,23]. Although not EFTR-specific, these findings transfer directly, as EFTR shares dissection demands (marking, incision, layer separation)[12].

Transferable surgical AI further validates real-time analysis for endoscopic/robotic workflows[24,25]. For TS, thoracic/robotic studies enable phase recognition, instrument tracking, and standardization, though esophageal SEL data lag[26,27]. Table 2[1,4,18] contrasts accuracies, highlighting endoscopic AI’s edge for EFTR (92.1%, 30 milliseconds latency) over TS planning tools (82%, 50 milliseconds). Thus, AI workflow tools emerge as enablers for EFTR/TS efficiency, pending SEL-specific validation[19].

Table 2 Published artificial intelligence performance metrics.

Application	Study focus	Accuracy (%)	Latency
EUS lesion classification	Esophageal SELs	92
ESD phase recognition	Esophageal	92.1	30 milliseconds
ESD phase recognition	Esophageal	84.3	40 milliseconds
Robotic TS phase recognition	Thoracic surgery	82	50 milliseconds

Artificial intelligence performance metrics from multicenter validation studies. EUS: Endoscopic ultrasound; ESD: Endoscopic submucosal dissection; SEL: Subepithelial lesions; TS: Thoracoscopic surgery.

ANATOMICAL ORIENTATION AND COMPLICATIONS

AI-assisted anatomical recognition proves valuable for EFTR, where precise layer identification underpins safe full-thickness excision[21]. Esophageal endoscopy studies demonstrate real-time segmentation of mucosa/muscularis and navigation support during dissection-directly transferable to EFTR margins and closure[26,28].

In TS, thoracic/robotic AI excels at vessel/nerve overlays and boundary guidance, aiding azygos mobilization and vagus preservation[9,29]. AI intraoperative navigation further protects recurrent laryngeal nerves in esophageal procedures. Complication prediction remains prospective: EFTR AI may flag stricture/perforation/closure risks; TS tools detect bleeding/chylothorax[27,30].

FUTURE DIRECTIONS

Future research must advance from feasibility to prospective SEL-specific validation in EFTR/TS cohorts. Key priorities include multicenter trials with standardized annotations, external datasets, and patient-centered endpoints (resection quality, adverse events, recovery)[19,31]. Hybrid systems that combine endoscopic computer-aided detection with thoracic or robotic augmented reality deserve particular attention[26]. These platforms may help unify lesion characterization, intraoperative navigation, and complication prediction within one workflow. However, their eventual clinical value will depend on rigorous comparative trials with patient-centered endpoints such as resection quality, adverse event rates, recovery time, and learning-curve effects[1,2,32].

LIMITATIONS

This review is narrative rather than systematic, so study selection and emphasis may reflect author judgment and the availability of published evidence. The direct evidence base for AI in complex esophageal SELs treated by EFTR or TS remains small, requiring careful use of adjacent and transferable studies to support interpretation. Heterogeneity in imaging protocols, AI architectures, and outcomes also limits direct comparison across studies and prevents meta-analysis.

Another limitation is that several AI performance metrics derive from non-SEL or non-esophageal settings, so they should be understood as supportive rather than definitive for the target clinical problem. For these reasons, the manuscript should be read as a structured synthesis of emerging evidence rather than a statement of established practice.

CONCLUSION

AI is best understood as an adjunctive technology for complex esophageal SEL management rather than a replacement for expert judgment in EFTR or TS. Current evidence supports three near-term roles: Improving lesion characterization with AI-EUS, assisting workflow recognition and anatomical orientation, and helping predict procedural complications. Endoscopic AI appears most immediately ready for real-time support, while thoracic and robotic AI is more advanced for planning and navigation. The major barrier to clinical translation remains the lack of large, SEL-specific, multicenter validation studies. Future work should therefore prioritize prospective trials, federated datasets, and hybrid AI systems that directly measure patient outcomes and surgical efficiency. Until then, AI should be presented as a promising support tool, not a proven standard of care.