Wision A.I. Achieves High Sensitivity and Specificity in Preclinical Validation of AI-Powered Early Diagnosis for Esophageal Squamous Cell Carcinoma

In January 2020, Gastrointestinal Endoscopy (GIE) [IF=7.229], a top-tier journal in the field of digestive endoscopy, published a preclinical validation paper by Professor Hu Bing’s team from West China Hospital and Wision AI on AI-based detection of precancerous lesions in esophageal squamous cell carcinoma (ESCC).^【1】Based on the information disclosed in the paper, we observed the high sensitivity and high specificity demonstrated by Wision AI’s independently developed Computer-Aided Diagnosis (CAD) system in the auxiliary diagnosis of precancerous lesions and early esophageal squamous cell carcinoma (ESCC).

According to preclinical study data, CAD was validated using endoscopic image and video datasets, ultimately achieving a sensitivity of 98.04% and a specificity of 95.03% in preclinical settings. It holds significant potential to assist endoscopists in diagnosing precancerous lesions of esophageal squamous cell carcinoma (ESCC) in the future.

This preclinical study not only validates the performance of Wision AI’s algorithm in identifying esophageal precancerous lesions, but also marks the first AI validation in the field of esophageal cancer that fully aligns with clinical diagnostic and treatment pathways.

Previously, CAD research on esophageal cancer was confined to the identification or classification of advanced-stage cancers under white-light endoscopy. However, even under conventional white-light endoscopy, the visual features of advanced esophageal cancer are highly conspicuous, resulting in inherently low rates of misdiagnosis and missed diagnosis by physicians. Consequently, the clinical utility of such CAD systems has been very limited.

The true clinical challenge lies in the early detection of esophageal cancer and its precancerous lesions. According to Japanese clinical guidelines, it has been demonstrated that the diagnosis of precancerous esophageal lesions relies solely on narrow-band imaging (NBI) magnifying endoscopy. However, due to the complexity and subtlety of the associated features, the misdiagnosis rate among inexperienced physicians approaches 50%.^【2】Therefore, Wision AI’s technology aims to address the AI-assisted detection of precancerous lesions under narrow-band imaging endoscopy.

As a preclinical validation aimed at verifying the sensitivity and specificity of the technology on a per-frame basis, the quality of the preclinical validation results and the rigor of the testing methodology indicate whether this technology can achieve favorable outcomes in real-world clinical applications.

In both academia and industry, some institutions attempt to artificially fabricate favorable fitting results by violating the independence between training and test samples—reducing the test set to merely 1/10 or 1/100 of the training set and selecting specific populations as test subjects—thereby publishing preclinical validation papers on AI research while completely ignoring the risk of overfitting in deep learning. This practice is corroborated by a statistical study conducted by Korean scholars, which showed that among 516 medical imaging AI research papers published online from January to August 2018, only 6% maintained complete independence between their training and test samples.^【3】！

As indicated in the preclinical validation paper published by Wision AI, the researchers prospectively and independently separated the training and validation datasets. The training dataset comprised 6,473 narrow-band imaging (NBI) images covering precancerous lesions, early-stage esophageal squamous cell carcinoma (ESCC), and benign lesions, which were provided by several specialized medical institutions, including the Endoscopy Center of West China Hospital (WCH) in Chengdu, China, and Jaswant Rai Specialty Hospital in Meerut, India.

Wision AI’s validation samples were divided into four datasets, evaluating the CAD system across multiple dimensions, including image and video validation. The validation data comprised a total of 175,536 images and video frames, approximately 27 times the size of the training dataset. According to the company’s validation results, with appropriate training data and techniques, the CAD system and its functionalities demonstrate significant potential in accurately pinpointing the locations of esophageal precancerous lesions and esophageal squamous cell carcinoma (ESCC).

Technically, Wision AI employs a single AI model to process lesion recognition under both standard and magnified Narrow Band Imaging (NBI). Standard and magnified NBI are both clinically significant for the diagnosis of precancerous lesions and esophageal squamous cell carcinoma (ESCC): brownish areas observed under standard NBI are key features of precancerous lesions and ESCC, while intrapapillary capillary loops (IPCLs) are the primary characteristics identified under magnified NBI. The Computer-Aided Diagnosis (CAD) system developed by Wision AI is designed to enable real-time, automatic diagnosis of precancerous lesions and early-stage ESCC under both standard and magnified NBI without requiring mode switching.

The ultimate goal of preclinical validation is to translate findings into clinical applications. However, because some preclinical studies fail to mitigate the risk of overfitting in deep learning, their results remain confined to preclinical data. Such occurrences are all too common:

In 2016, Google announced its preclinical study on AI-based diagnosis of diabetic retinopathy (DR).^【4】Preclinical studies have demonstrated that Google’s proprietary deep learning system can automatically detect diabetic retinopathy (DR), thereby mitigating the risk of irreversible blindness in hundreds of millions of patients with diabetes. However, this technology has failed to effectively advance to subsequent clinical trials.^【5】。

Coincidentally, in 2017, a Stanford paper published in Nature announced a new breakthrough: deep learning achieved expert-level performance in skin cancer diagnosis, enabling the system to perform automated skin cancer diagnosis for patients.^【6】. However, years have passed, and the much-anticipated follow-up on automated diagnosis of skin cancer has yet to appear, ultimately fading away without result.

Excellent preclinical research paves the way for subsequent clinical applications. In 2018, Wision AI published a preclinical paper titled “Development and validation of a deep learning algorithm for detection of polyps during colonoscopy” in the top academic journal Nature.^【7】, demonstrating that a machine learning algorithm can detect polyps in clinical colonoscopy in real time with high sensitivity and high specificity. The company has completed a prospective randomized controlled study of this system.

In an interview, Liu Jingjia, founder of Wision AI, also stated that a key component of the company’s strategy is to conduct rigorous trials on product performance and publish these studies in high-impact journals, thereby generating globally recognized, high-quality clinical evidence. Only by adhering to rigorous principles of evidence-based medicine can products effectively improve core clinical outcomes and gain widespread acceptance.

Esophageal cancer is one of the most common malignant tumors worldwide. As the predominant subtype of esophageal cancer, esophageal squamous cell carcinoma (ESCC) accounts for more than 90% of all esophageal cancer cases in China, with an overall 5-year survival rate of less than 20%. Therefore, early diagnosis of precancerous lesions and ESCC is crucial for favorable patient prognosis.

However, the early imaging features of esophageal squamous cell carcinoma (ESCC) are difficult to identify. When endoscopists with limited experience use narrow-band imaging (NBI), the sensitivity for detecting ESCC is only 53%. A recent study on missed diagnoses of esophageal cancer found that 6.4% of patients had negative endoscopic results within the three years preceding their diagnosis. Due to a shortage of well-trained endoscopists, particularly in rural or underdeveloped areas, equipping them with the capability to detect precancerous lesions and ESCC remains a significant challenge.

In recent years, significant progress has been made in computer-aided diagnosis (CAD) using artificial intelligence (AI) systems. Researchers have employed CAD systems to improve the diagnosis of various gastrointestinal lesions, such as colorectal polyps, gastric ulcers, Helicobacter pylori infection, and gastric cancer. Meanwhile, the application of CAD for the early diagnosis of esophageal squamous cell carcinoma (ESCC) has also garnered widespread attention.

In 2018, Horie et al. developed a deep learning model based on magnified narrow-band imaging (NBI) images to study the automatic classification of intrapapillary capillary loops (IPCLs), marking the first use of artificial intelligence for esophageal cancer detection, with a sensitivity of 98% and a positive predictive value of 40%. However, their study only tested static images, did not demonstrate the differences between standard and magnified settings, and did not employ real-time analysis; the research findings focused primarily on the classification rather than the detection of NBI images.

Wision AI’s computer-aided diagnosis (CAD) system for the automated detection of esophageal squamous cell carcinoma (ESCC) precancerous lesions leverages artificial intelligence to replicate the diagnostic expertise of skilled endoscopists, thereby effectively expanding screening capacity and improving the current landscape of early diagnosis for esophageal malignancies.

Previously, Wision AI had already developed mature AI-powered clinical decision support products, primarily applied in the field of early screening for colorectal cancer. These solutions have been implemented in clinical practice and jointly validated by numerous leading domestic and international hospitals and clinical academic authorities, including Harvard Medical School, effectively improving a core clinical metric in gastrointestinal early cancer screening: the detection rate of early-stage cancers and precancerous lesions.

Blooming within China, fragrant beyond its borders. The clinical trial data and conclusions of Wision AI have garnered significant attention and recognition from authoritative international institutions, with findings published in multiple top-tier international clinical journals. Wision AI has received numerous awards from the World Congress of Gastroenterology (WCOG), the American College of Gastroenterology (ACG) Annual Scientific Meeting, and United European Gastroenterology (UEG) Week. It has drawn high-level attention from leading academic authorities and regulatory bodies in Europe, the United States, and Japan specializing in early gastrointestinal cancer screening. The President of the American Gastroenterological Association hailed it as a “revolutionary innovation,” while the U.S. Food and Drug Administration (FDA) described it as a “game changer.” Its researchers were recognized for best authorship and papers by prestigious academic journals such as Gut and Gastrointestinal Endoscopy (GIE). Furthermore, Wision AI has earned unanimous acclaim from key opinion leaders in the international endoscopy community, including Professor Haruhiro Inoue, renowned as the “God of Endoscopy” in Japan. With great promise, Wision AI is poised to become the benchmark enterprise in the global field of AI for digestive endoscopy.

References:

【1】Guo, LinJie et al.  Real-time automated diagnosis of precancerous lesions and early esophageal squamous cell carcinoma using a deep learning model (with videos) Gastrointestinal Endoscopy, Volume 91, Issue 1, 41 - 51

【2】Ishihara R, Takeuchi Y, Chatani R, et al. Prospective evaluation of narrow-band imaging endoscopy for screening of esophageal squamous mucosal high-grade neoplasia in experienced and less experienced endoscopists. Dis Esophagus 2010;23:480-6.

【3】Kim DW, Jang HY, Kim KW, Shin Y, Park SH.   Design Characteristics of Studies Reporting the Performance of Artificial Intelligence Algorithms for Diagnostic Analysis of Medical Images: Results from Recently Published Papers.   Korean J Radiol. 2019 Mar;20(3):405-410.

【4】Gulshan, V. et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA 316, 2402–2410 (2016).

【5】https://www.wsj.com/articles/googles-effort-to-prevent-blindness-hits-roadblock-11548504004

【6】Esteva, A. et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature 542, 115–118 (2017).

【7】Wang P, Xiao X, Glissen Brown JR, et al. Development and validation of a deeplearning

algorithm for the detection of polyps during colonoscopy. Nat Biomed Eng 2018;2:741–8.