Artificial intelligence-aided optical biopsy improves the diagnosis of esophageal squamous neoplasm

Abstract

BACKGROUND

Early detection of esophageal squamous neoplasms (ESN) is essential for improving patient prognosis. Optical diagnosis of ESN remains challenging. Probe-based confocal laser endomicroscopy (pCLE) enables accurate in vivo histological observation and optical biopsy of ESN. However, interpretation of pCLE images requires histopathological expertise and extensive training. Artificial intelligence (AI) has been widely applied in digestive endoscopy; however, AI for pCLE diagnosis of ESN has not been reported.

AIM

To develop a pCLE computer-aided diagnostic system for ESN and assess its diagnostic performance and assistant efficiency for nonexpert endoscopists.

METHODS

The intelligent confocal laser endomicroscopy (iCLE) system consists of image recognition (based on inception-ResNet V2), video diagnosis, and quality judgment modules. This system was developed using pCLE images and videos and evaluated through image and prospective video recognition tests. Patients between June 2020 and January 2023 were prospectively enrolled. Expert and non-expert endoscopists and the iCLE independently performed diagnoses for pCLE videos, with histopathology as the gold standard. Thereafter, the non-expert endoscopists performed a second assessment with iCLE assistance.

RESULTS

A total of 25056 images from 2803 patients were selected for iCLE training and validation. Another 2442 images from 226 patients were used for testing. iCLE achieved a high accuracy of 98.3%, sensitivity of 95.3% and specificity of 98.8% for diagnosing ESN images. A total of 2581 patients underwent upper gastrointestinal pCLE examination and were prospectively screened; 54 patients with suspected ESN were enrolled. Overall, 187 videos from 67 lesions were assessed by iCLE, three nonexpert and three expert endoscopists. iCLE achieved a high accuracy, sensitivity and specificity of 90.9%, 92.0%, and 90.2%, respectively. Compared to experts, iCLE showed significantly higher sensitivity (92.0% vs 80.4%; P < 0.001) and negative predictive value (94.4% vs 87.7%; P = 0.003). With iCLE assistance, nonexpert endoscopists showed significant improvements in accuracy (from 83.6% to 88.6%) and sensitivity (from 76.0% to 89.8%).

CONCLUSION

iCLE system demonstrated high diagnostic performance for ESN. It can assist nonexpert endoscopists in improving the diagnostic efficiency of pCLE for ESN and has the potential for reducing unnecessary biopsies.

Key Words: Esophageal squamous neoplasm; Probe-based confocal laser endomicroscopy; Optical biopsy; Artificial intelligence; Computer aided diagnosis

Core Tip: The optical diagnosis of esophageal squamous neoplasms (ESN) is challenging. Probe-based confocal laser endomicroscopy (pCLE) enables the optical biopsy of ESN; however, its application is affected by the difficulty of image interpretation. This study developed the first pCLE computer-aided diagnostic system (intelligent confocal laser endomicroscopy) for ESN, enabling real-time diagnosis of pCLE videos. Intelligent confocal laser endomicroscopy was evaluated in a prospective study and compared with endoscopists with different expertise levels. It demonstrated higher sensitivity than both nonexpert and expert endoscopists. Additionally, it assists nonexperts in significantly improving diagnostic accuracy and sensitivity. This system has the potential to assist endoscopists in the application of pCLE and reduce unnecessary biopsies.

INTRODUCTION

Esophageal cancer is a major global health issue, ranking as the 11^th most commonly diagnosed cancer and 7^th leading cause of cancer-related mortality worldwide[1]. Epidemiological studies have shown a high disease burden, with more than 85% of global esophageal cancer cases attributed to squamous cell carcinomas (SCC), especially in central and eastern Asia. Nearly half of all cases worldwide occur in China[2,3]. Early detection and treatment of esophageal squamous neoplasms (ESN) are essential for improving patient prognosis and reducing the disease burden. Because the risk of lymph node metastases in superficial esophageal cancer correlates with the depth of lesion invasion, early identification of ESN enables patients to receive minimally invasive treatments, such as endoscopic submucosal dissection (ESD), rather than surgical interventions[4].

Endoscopy, in conjunction with biopsy sampling, is the primary diagnostic method for ESN. The identification of ESN using conventional white-light endoscopy (WLE) is challenging owing to their subtle appearance, leading to a low diagnostic sensitivity of 55%[5]. Various endoscopic techniques such as Lugol’s iodine chromoendoscopy (LCE), image-enhanced endoscopy (IEE), magnifying endoscopy (ME), and probe-based confocal laser endomicroscopy (pCLE) have been used to improve the detection and diagnostic accuracy for ESN. LCE is economical, widely used, and has high sensitivity (80%-100%) for diagnosing ESN. However, its specificity is low (64%-94%) because areas of neoplasms, inflammation, and columnar epithelium can all appear unstained. In addition, iodine spraying may cause patient discomfort such as heartburn. IEE, including narrow-band imaging (NBI), has shown equivalent sensitivity (82%-88%) as that of LCE, with higher specificity (75%-95%)[4]. However, biopsy and histopathological examinations are still required to differentiate ESN from benign lesions after LCE or NBI, especially in patients with multiple esophageal lesions. To achieve a precise optical diagnosis and minimize biopsy sampling, IEE can be combined with ME (ME-IEE) or pCLE.

pCLE enables real-time, in vivo histological observation of the gastrointestinal mucosa at the cellular and microvascular levels. This approach aims to overcome the inherent limitations of traditional endoscopic sampling techniques using optical biopsy combined with targeted biopsy[6]. The mini-probe of the pCLE can be inserted into the working channel of other endoscopes, such as a high-definition endoscope with NBI, allowing microscopic imaging via direct mucosal contact. Previous studies have demonstrated that pCLE is highly effective for the diagnosis of ESN, with both high sensitivity (94%-95.7%) and specificity (90%)[4,6-8]. Notably, pCLE can accurately distinguish benign from malignant Lugol-unstained esophageal lesions in patients with head and neck cancers[7]. In addition, pCLE can be efficiently used in combination with IEE to reduce the number of biopsy specimens required for diagnosis[8]. However, the interpretation of pCLE images requires a substantial knowledge of pathology and professional training, which limits its application and promotion[9].

Despite the variety of endoscopic equipment available, a proportion of lesions remain misdiagnosed owing to variability in endoscopists’ expertise, visual fatigue, and other factors[10]. Artificial intelligence (AI), which can detect suspicious lesions and improve diagnostic accuracy, has been widely applied in endoscopic diagnosis. Previous studies have shown that AI systems can accurately diagnose superficial esophageal SCC and precancerous lesions using WLE and non-magnified NBI, thereby improving the detection rates[11,12]. In addition, AI can recognize the microvascular classification of ESN and determine invasion depth using ME-IEE[13,14]. However, to our knowledge, no AI program has been developed to diagnose ESN using pCLE.

In this study, our team developed a pCLE computer-aided diagnostic system, termed intelligent confocal laser endomicroscopy (iCLE), for the diagnosis of ESN based on pCLE. This study aimed to compare the diagnostic performance of AI with that of endoscopists and to evaluate its utility in assisting nonexpert endoscopists.

MATERIALS AND METHODS

Study design

This prospective diagnostic study was conducted at the Qilu Hospital of Shandong University (Jinan, Shandong Province, China). The iCLE system was developed using deep-learning networks. pCLE images and videos were retrospectively collected from the pCLE database of the Qilu Hospital for the development, validation, and image testing of the iCLE system. A prospective study was conducted to evaluate the diagnostic performance of iCLE for pCLE videos and to compare it with that of endoscopists with varying levels of expertise. In addition, the assistant efficiency of iCLE for non-expert endoscopists was also assessed.

This study was approved by the Ethics Committee of Qilu Hospital of Shandong University and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all patients. The study was registered at https://clinicaltrials.gov/ (No. NCT04136236).

Collection and preprocessing of images

pCLE images and videos obtained between April 2014 and May 2020 were retrospectively collected for the development of the iCLE system. Data were obtained from consecutive patients who underwent pCLE examinations. Videos of the esophagus were reviewed and extracted into single-frame images. A deduplication software was developed to eliminate highly similar frames by comparing the Hamming distances between them. A workflow diagram depicting this process is shown in Supplementary Figure 1. This approach ensures that the selected images from the same lesion are representative and non-redundant. Images with typical pCLE features of SCC, high-grade intraepithelial neoplasia (HGIN), low-grade intraepithelial neoplasia (LGIN), inflammation, normal squamous epithelium, and unqualified images (images with invisible tissue structure, too dark or too bright vision, or artifacts) were selected and labeled. SCC, HGIN, and LGIN were classified as neoplasms, whereas inflammation and normal squamous epithelium were classified as non-neoplasms. Some videos were edited into video segments (each short video contained images of the probe touching the mucosa) to train the iCLE for video recognition. The most severe histological abnormality in each video was considered the final diagnosis.

The diagnostic criteria for pCLE were based on the surface maturation scoring (SMS) system[8]. This system evaluates the features of surface maturation in the esophageal squamous epithelium, including the existence of halos, gradients, polarity, and the compass effect. The absence of all of these features (SMS score = 0) was considered a neoplasm. All images and videos were reviewed and marked by two experienced endoscopists, each with experience in over 100 pCLE procedures, and verified by another senior expert endoscopist. Histological diagnosis served as the gold standard for labeling. If the diagnosis of one image or video was discrepant, the three endoscopists discussed the findings to reach a consensus; if consensus was unattainable, the image or video was discarded.

The images from the different patients were then partitioned into three distinct sets: Training, validation, and test sets. This approach effectively prevented data leakage by ensuring that images from the same patient were contained exclusively within a single dataset. The labeled pCLE images underwent further preprocessing, including image cropping, resizing, and whitening, to eliminate interference factors and enhance image features. Subsequently, the images were converted into grayscale images with a size (299 × 299 × 1) and input into the model.

Development of the AI model

The iCLE system was constructed using neural networks and included three submodules: Image recognition network (Supplementary Figure 2), video diagnosis network, and quality judgment network. The performance of the various deep neural networks, including VGG, GoogleLeNet, RestNet, and inception-ResNet V2 were tested for pCLE image recognition efficiency. Among these, inception-ResNet V2 demonstrated the best performance, and was therefore selected as the final model of the image recognition network. The model was trained and validated using the TensorFlow 1.9.0 framework. During training, 32 images were randomly selected from the training set (batch size = 32) for each training iteration. After each epoch, the model was validated using the validation set. The training was stopped if the validation accuracy did not improve over 20 consecutive epochs, with the maximum number of epochs set to 200. Training was performed using gradient descent, with the Adam optimizer employed to optimize the network parameters. The initial learning rate was set to 0.001 and the learning rate adjustment strategy was based on Adam’s dynamic learning rate mechanism. This network used a pCLE image as the input and generated a three-dimensional feature vector. The feature vector was then processed using Softmax and converted into probabilities for the three diagnostic categories. Subsequently, the image recognition network output the diagnosis with the highest the probability, along with its corresponding probability value. A fully connected neural network was used to diagnose the videos. The AI-derived diagnosis and diagnostic probability for each image in the video were used as the input, and the diagnosis for the video served as the output. The model was trained using video segments. The quality judgment network was a two-layer convolutional neural network that can distinguish between qualified and unqualified images. The model was trained using both qualified and unqualified images. A workflow diagram for the development of the iCLE is shown in Figure 1. The real-time diagnosis of iCLE was displayed on a screen (Figure 2). The diagnostic performance for the model is shown in Figure 3 and Video 1.

Figure 1 Workflow diagram for the development of intelligent confocal laser endomicroscopy. iCLE: Intelligent confocal laser endomicroscopy; pCLE: Probe-based confocal laser endomicroscopy.

Figure 2 Work interface of intelligent confocal laser endomicroscopy. The intelligent confocal laser endomicroscopy system could show the images of probe-based confocal laser endomicroscopy with real-time artificial intelligence diagnosis, white-light endoscopy or narrow-band imaging, and the artificial intelligence diagnosis for each frame of the video.

Figure 3 Diagnosis of intelligent confocal laser endomicroscopy for esophageal lesions. A: A brownish lesion observed on narrow-band imaging (NBI); B: Probe-based confocal laser endomicroscopy (pCLE) image of inflammatory epithelium; Intelligent confocal laser endomicroscopy (iCLE) shows the diagnosis of non-neoplasm; C: The lesion shows Lugol-voiding on Lugol’s iodine chromoendoscopy (LCE) after the pCLE procedure; D: Another brownish lesion observed on NBI; E: PCLE image of esophageal squamous neoplasms with the squamous maturation score was 0; ICLE shows the diagnosis of neoplasm; F: Lugol-voiding observed on LCE.

Validation of iCLE

The diagnostic capability of iCLE was first evaluated using still pCLE images. Images from the test set were analyzed using a computer equipped with the AI. Then a prospective study was conducted.

Patients

From June 2020 to January 2023, patients who underwent upper gastrointestinal pCLE at the Qilu Hospital of Shandong University for screening, surveillance, or evaluation of gastrointestinal symptoms were included in the study. Patients aged 18-80 years who provided informed consent were included in the study. Exclusion criteria included advanced esophageal SSC or esophageal stenosis, the absence of suspicious ESN lesions on WLE or IEE, known allergy to fluorescein sodium, coagulopathy or impaired renal function, and pregnancy or breastfeeding.

Sample size calculation

The sample size was calculated using PASS15 (NCSS, LLC, Kaysville, MD, United States). We assumed that the sensitivity of non-expert endoscopists for pCLE diagnosis of ESN was 80% and would improve to 90% with iCLE assistance. A sample size of 137 was required to achieve 90% statistical power to detect a difference, using a two-sided exact test with a significance level (alpha) of 0.05. If the dropout rate was 20%, then at least 172 videos should be recognized.

Endoscopic procedure and histological evaluation

Expert endoscopists, each with experience in more than 100 pCLE examinations, conducted all endoscopic procedures. These were performed using either a GIF-HQ290 endoscope with an Evis Lucera Elite CV-290 system (Olympus, Tokyo, Japan) or an EG29-i10 endoscope with an EPK-i7000 image processor (Pentax, Tokyo, Japan). A Cellvizio probe (Cellvizio, Paris, France) was used to perform the pCLE. Following routine preparation for anesthetic gastroscopy, the patients underwent upper gastrointestinal endoscopy under general anesthesia and electrocardiographic monitoring. The esophagus was observed using both WLE and IEE, and suspicious lesions were recorded along with their macroscopic characteristics, location, and size. All the IEE observations were performed without magnification. Suspicious esophageal lesions were examined using pCLE for further differentiation. Fluorescein sodium (0.02 mL/kg, 10%) was administered intravenously and the pCLE probe was inserted through the working channel of the endoscope. Multipoint scanning was performed from the periphery to the center of the lesion, and the number of observation points was determined based on the lesion size. At each observation point, the probe remained in place until a recognizable, stable video was obtained. Endomicroscopic videos of different lesions were stored separately for further analysis. Finally, targeted biopsies were obtained from the positions with the most histological abnormalities identified using pCLE. All the specimens were sent for histopathological examination and evaluated by experienced pathologists. Diagnosis was based on the Vienna classification[15].

Diagnosis of pCLE videos

After examination, the pCLE videos were divided into recognizable video segments (containing sufficient pCLE images for diagnosis) and anonymized. For large lesions, pCLE videos were split into multiple non-overlapping video segments. The iCLE and endoscopists (three experts and three non-experts) classified each video as either neoplasm or non-neoplasm. The pCLE videos were displayed on a computer in a predetermined order for each endoscopist to make a diagnosis, whereas the iCLE processed the videos on another computer. Endoscopists were blinded to the endoscopic and pathological diagnoses. Endoscopists with experience in at least 100 pCLE procedures were defined as experts; non-expert endoscopists were those who could perform routine gastroscopy and had learned about pCLE, but had performed fewer than 100 procedures. After a 2-week washout period, non-expert endoscopists examined the videos again, which presented the AI diagnoses. Histological diagnosis was considered the gold standard. The workflow for evaluating the iCLE system is shown in Figure 4. Information on the endoscopists involved in the diagnosis is shown in Supplementary Table 1.

Figure 4 Workflow diagram for the prospective validation of intelligent confocal laser endomicroscopy. ESN: Esophageal squamous neoplasms; iCLE: Intelligent confocal laser endomicroscopy; pCLE: Probe-based confocal laser endomicroscopy.

Outcomes

The primary outcome measure was the diagnostic performance of iCLE in a prospective study, assessed using diagnostic accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) for ESN classification. The secondary outcome measure was the comparison of the diagnostic performances of iCLE, expert, and non-expert endoscopists, as well as the evaluation of the assistance provided by iCLE to non-expert endoscopists.

Statistical analysis

McNemar’s test was used to compare the diagnostic performances of iCLE, endoscopists with different expertise levels, and AI-assisted non-experts. Statistical significance was defined as a two-sided P value < 0.05. All statistical analyses were performed using the SPSS version 25 software (IBM Corp., Armonk, NY, United States). All authors have access to the study data and have reviewed and approved the final manuscript.

RESULTS

A total of 3029 patient pCLE images and videos from Qilu Hospital of Shandong University were retrospectively collected. After the selection and deduplication process, a total of 22972 images (3220 neoplasm, 12729 non-neoplasm, and 7023 unqualified images) from 2564 patients were allocated to the training set. Additionally, 2084 images (341 neoplasm, 1002 non-neoplasm, and 741 unqualified images) from 239 patients were assigned to the validation set. Another 2442 images (358 neoplasm, 1362 non-neoplasm, and 722 unqualified images) from 226 patients were used as the test set. A total of 85 video segments were collected to train the video recognition ability. Between June 2020 and January 2023, 2581 patients underwent upper gastrointestinal pCLE examination and were prospectively screened; of these, 54 patients with suspected ESN were enrolled according to the inclusion and exclusion criteria. Eventually, 67 lesions from 54 patients were detected, and 187 videos were used for video testing. No adverse events were reported. The baseline characteristics of the patients and their lesions are shown in Table 1.

Table 1 Baseline characteristics of patients and lesions, n (%).

Baseline characteristics	n
Patients	54
Male	39 (72.2)
Female	15 (27.8)
Age, mean (range), years	59.3 (38-77)
Lesions	67
Lesion size, mean, cm	1.26
Location
Upper esophagus	5
Middle esophagus	35
Lower esophagus	27
Macroscopic type
I	3
0-IIa	3
0-IIb	58
0-IIa + IIc	3
Histological diagnosis
Esophageal squamous cell carcinoma	8
HGIN	9
LGIN	8
Inflammatory/hyperplastic lesions	42

HGIN: High-grade intraepithelial neoplasia; LGIN: Low-grade intraepithelial neoplasia.

The diagnostic performance of iCLE was satisfactory. iCLE achieved a high accuracy of 98.3% (95% confidence interval, 97.8%-98.8%), sensitivity of 95.3% (93.0%-97.5%), specificity of 98.8% (98.4%-99.3%), PPV of 93.4% (90.9%-96.0%), and NPV of 99.2% (98.8%-99.6%) for diagnosing ESN images. The receiver operating characteristic curve (Figure 5) showed a high area under the curve (0.9878-0.9959). The decision curve analysis curve is shown in Supplementary Figure 3. Table 2 shows the diagnostic performance of the pCLE videos. The diagnostic measurements of iCLE and experts were significantly superior to those of nonexperts. iCLE achieved 90.9% (88.5%-93.3%) accuracy, 92.0% (88.4%-95.6%) sensitivity, 90.2% (87.0%-93.4%) specificity, 86.3% (81.9%-90.6%) PPV, and 94.4% (91.9%-96.9%) NPV for ESN. Compared with experts, iCLE showed significantly higher sensitivity (92.0% vs 80.4%; P < 0.001) and NPV (94.4% vs 87.7%; P = 0.003), with similar accuracy, specificity, and PPV. With the assistance of iCLE, nonexperts had significant improvement in their accuracy from 83.6% (80.5%-86.7%) to 88.6% (86.0%-91.2%) (P = 0.016), sensitivity from 76.0% (70.4%-81.6%) to 89.8% (85.8%-93.8%) (P < 0.001), and NPV from 84.7% (80.9%-88.4%) to 92.8% (89.9%-95.6%) (P = 0.001). In addition, they achieved higher sensitivity (89.8% vs 80.4%; P = 0.005) and NPV (92.8% vs 87.7%; P = 0.028) than experts, with similar accuracy and PPV, but inferior specificity (87.8% vs 93.5 %; P = 0.012).

Figure 5 Receiver operating characteristic curve for intelligent confocal laser endomicroscopy in the diagnosis of probe-based confocal laser endomicroscopy images. ROC: Receiver operating characteristic curve; AUC: Area under curve.

Table 2 Diagnostic performance for probe-based confocal laser endomicroscopy videos.

Parameter	iCLE, % (95%CI)	Expert, % (95%CI)	NE, % (95%CI)	NE with iCLE, % (95%CI)	P value	Expert vs NE	iCLE vs NE	iCLE vs expert	NE with iCLE vs NE	NE with iCLE vs expert
Accuracy	90.9 (88.5-93.3)	88.2 (85.6-90.9)	83.6 (80.5-86.7)	88.6 (86.0-91.2)		0.026a	< 0.001a	0.143	0.016a	0.863
Sensitivity	92.0 (88.4-95.6)	80.4 (75.2-85.7)	76.0 (70.4-81.6)	89.8 (85.8-93.8)		0.254	< 0.001a	< 0.001a	< 0.001a	0.005a
Specificity	90.2 (87.0-93.4)	93.5 (90.8-96.1)	88.7 (85.3-92.1)	87.8 (84.3-91.3)		0.031a	0.532	0.121	0.718	0.012a
PPV	86.3 (81.9-90.6)	89.2 (84.8-93.5)	81.8 (76.5-87.1)	83.1 (78.4-87.9)		0.035a	0.199	0.354	0.718	0.068
NPV	94.4 (91.9-96.9)	87.7 (84.3-91.1)	84.7 (80.9-88.4)	92.8 (89.9-95.6)		0.238	< 0.001a	0.003a	0.001a	0.028a

^aP < 0.05.

NE: Nonexpert; iCLE: Intelligent confocal laser endomicroscopy; CI: Confidence interval; PPV: Positive predictive value; NPV: Negative predictive value.

DISCUSSION

In this study, we developed a computer-aided pCLE diagnostic system for ESN diagnosis and evaluated its effectiveness using images and prospective video validation. iCLE achieved high diagnostic performance in both tests and was superior to expert and non-expert endoscopists in video diagnosis. AI assistance significantly improved the diagnostic efficiency of non-expert endoscopists and achieved higher sensitivity than experts. To the best of our knowledge, iCLE is the first AI program for pCLE that enables the in vivo histological evaluation of esophageal squamous epithelial lesions.

The LCE and IEE are the most commonly used techniques for detecting ESN. LCE has high sensitivity but is limited by the low specificity and discomfort associated with the procedure. IEE has higher specificity and similar sensitivity compared to LCE and can be combined with other techniques for optical diagnosis. Our previous study demonstrated that IEE combined with pCLE is an effective diagnostic strategy[8]. pCLE provides real-time histological assessments by visualizing intraepillary capillary and squamous epithelial cells. While IEE identifies suspicious lesions as red flags, pCLE differentiates and guides targeted biopsies, which can improve both detection rates and diagnostic accuracy of ESN. However, ESN diagnosis is based on the SMS system and requires endoscopists to acquire histopathological knowledge and experience. The skills and expertise of endoscopists vary considerably, and interobserver agreement for pCLE diagnosis between non-expert and specialist endoscopists is variable. It is difficult for non-expert endoscopists to achieve sufficient diagnostic performance. As shown in this study, the diagnostic accuracy, specificity, and PPV of non-expert endoscopists were significantly lower than those of the expert endoscopists. The subjective nature of pCLE diagnosis by non-expert endoscopists compromises the reliability of optical biopsies. Therefore, the development of computer-based algorithms for image interpretation is urgently required to improve CLE technology.

Although several effective AI systems have been developed to assist in the detection of ESN using WLE, IEE, or ME-IEE, iCLE represents the first AI system designed for pCLE-based diagnosis of ESN. The iCLE system achieved superior diagnostic performance compared to endoscopists with different expertise levels and demonstrated significantly higher sensitivity and NPV than both non-expert and expert endoscopists. Endoscopists were more cautious of diagnosing neoplastic lesions without the presentation of WLE and IEE images. Additionally, some LGIN lesions were diagnosed as inflammatory lesions by the endoscopists, leading to a lower sensitivity among them. The experts had the highest specificity of all groups, although the difference was not statistically significant compared to iCLE. Therefore, this does not mean that iCLE can completely surpass experts in routine examinations because experts can synthesize the macroscopic and microscopic manifestations of lesions using IEE and pCLE to make comprehensive judgments. However, in this study, AI achieved the best results in evaluating the diagnostic performance of the pCLE videos. Moreover, with the assistance of iCLE, non-expert endoscopists significantly improved their diagnostic performance, achieving a diagnostic accuracy similar to that of experts and a higher sensitivity. Thus, iCLE is an effective auxiliary diagnostic tool to assist endoscopists in making precise real-time diagnoses of ESN.

Accurate biopsy acquisition is essential for determining the appropriate treatment or follow-up strategies. Non-targeted biopsies may result in the incorrect assessment and management of neoplastic lesions, potentially affecting patient prognosis. In routine examinations, although LCE or IEE are used to detect esophageal lesions, many biopsies are performed to distinguish between benign and malignant lesions, especially in cases of multicentric dysplasia, which is characterized by multiple Lugol-voiding lesions (LVLs) observed on LCE. Identifying the most severe lesions in patients with LVLs remains challenging. Generally, endoscopists tend to perform multiple biopsies to ensure histopathological accuracy. However, redundant biopsies increase the cost to patients and the workload of pathologists. In addition, it may lead to submucosal fibrosis of the lesion, complicating ESD and increasing the risk of muscle injury. As demonstrated by the Asian consensus for the diagnosis of upper gastrointestinal neoplasms, biopsy can be avoided if the endoscopic diagnosis achieves high accuracy for benign lesions and a high NPV for neoplasms[16]. iCLE can reduce the difficulty of pCLE analysis and provide objective referential diagnoses for non-expert endoscopists. Considering the high NPV of iCLE (94.4%), biopsies may not be required for non-neoplastic lesions diagnosed using AI, unless endoscopists have concerns. Targeted biopsies should be performed at sites where the AI indicates a neoplasm. Therefore, iCLE has the potential to improve biopsy accuracy and prevent misdiagnoses. It can also reduce the number of biopsies required for diagnosis and the associated costs of histopathological examinations, thereby improving the cost-effectiveness of gastroscopy. Considering the high sensitivity of iCLE, it may be helpful for determining the lateral margins for ESD, particularly in patients with LVLs.

This study has several strengths. First, the sample size of the training set was sufficiently large, encompassing data from 2564 patients. This extensive dataset allowed the AI to comprehensively learn the characteristics of pCLE images, guaranteeing its generalizability. Second, we developed a specialized software to facilitate the deduplication of pCLE images. Images in the test set collected from the same lesion may be highly similar, leading to overfitting of the neural network. Additionally, as it is difficult for humans to objectively assess the degree of resemblance between images, all images collected were selected using the deduplication software and highly similar images were discarded, which decreased selection bias and made the study more objective and convincing. Third, the iCLE was trained to recognize unqualified images and videos. The pCLE examinations may produce unqualified images, particularly when performed by non-expert endoscopists, as rapid movements during image acquisition create artifacts. These unqualified images may lead to misdiagnosis. We specifically designed a neural network in the iCLE to identify unqualified images, trained with a variety of such images and video segments. iCLE does not provide a diagnosis when recognizing unqualified images, which decreases their impact on the diagnosis.

This study has some limitations. Firstly, the diagnoses of the pCLE operators in this prospective study were not included in the comparison. As the operators had access to the WLE and IEE characteristics of the lesions, their diagnosis based on pCLE may have been inevitably affected, potentially leading to an overestimation of diagnostic performance. Second, the patients enrolled consisted of a general outpatient population rather than a high-risk population for ESN; as a result, limited lesions were detected despite the large number of patients undergoing the examination. Third, all the pCLE procedures were performed by experienced endoscopists. Although iCLE can identify unqualified images and videos, its diagnostic efficacy should be further assessed when general endoscopists perform pCLE. Finally, this was a single-center study; therefore, iCLE should be further evaluated in a multicenter trial.

CONCLUSION

We developed the first pCLE computer-aided diagnostic system for ESN. iCLE can achieve high-level diagnostic performance, assist non-expert endoscopists in the application of pCLE, and improve diagnostic effectiveness. iCLE has the potential to promote translation from physical biopsies to optical biopsies combined with targeted biopsies, and reduce unnecessary biopsies.