Introduction

AI diagnosis is used in several medical fields; however, machine learning using deep learning methods is particularly effective in diagnostic imaging and shows high diagnostic capability even in the field of endoscopy. In Japan, endoscopes manufactured by Companies A and B are generally used, and there is a possibility that differences in the image quality will affect AI diagnosis. However, there is no report of studies comparing the AI diagnostic capabilities on images taken by endoscopes of two companies.

Aims & Methods

In endoscopic AI diagnosis of early gastric cancer, the AI diagnostic capabilities were compared using images obtained by endoscopes of Companies A and B for the purpose of clarifying whether there is a need to consider the differences between the endoscopes used. For 30 lesions of 27 patients who underwent endoscopic submucosal dissection for early gastric cancer from April 2019 to March 2022 at our hospital, 356 and 338 endoscopic images of Companies A and B were the target of analysis, respectively.The images were white light (WLI), chromoendoscopy (indigo carmine) and image-enhanced endoscopy, Company A used narrow band imaging (NBI), and Company B used blue laser imaging (BLI).The images were analyzed by AI Medical Systems Inc. using an AI diagnostic system constructed by having UNET learn images taken by either Company A or B's endoscopes as training images. The diagnostic capability was assessed by comparing the specialist-marked correct range of early gastric cancer with an AI-diagnosed range of early gastric cancer, and using a DiceScore of 1.0 for a complete match and 0.0 for no match based on the area of the match.

Results

There was no significant difference in the DiceScore between the endoscopes of both companies in WLI; however, in indigo, Company B scored significantly higher than Company A. There was no significant difference between NBI and BLI. A stratification analysis by macroscopic examination of the lesions showed that in 0-IIc, Company A scored significantly higher in WLI than Company B, while in indigo, Company B scored significantly higher than Company A. In 0-IIa+IIc, Company B scored significantly higher than Company A in WLI; however, there was no difference between the two companies in indigo. In comparison by macroscopic examination of the lesions, 0-IIa+IIc lesions scored significantly higher than 0-IIc lesions in indigo for Company A, and in WLI for Company B. When the lesions were divided into a maximum diameter of <2 cm, 2–3 cm, and ≥3 cm, lesions ≥3 cm scored lower than lesions in the other groups.

Conclusion

Endoscopic images of Company B tended to have higher sharpness and stronger bluish color compared to endoscopic images of Company A. This might be the reason why the diagnostic capability improved after indigo-carmine scattering. Therefore, proactive scattering of indigo carmine was considered preferable when observing with the Company B endoscope. Presumably, the reason 0-IIa+IIc lesions had higher diagnostic performance than 0-IIc lesions was due to differences in the elevation of the lesions. The Company A endoscope was considered significantly superior for 0-IIc lesions, which are easily overlooked in clinical practice. The complexity of the morphology of the lesion increases in large lesions, which widens the gap between the AI diagnostic range and correct range. There are lesions wherein the AI diagnostic performance increased with both endoscopes of Company A and that of Company B. Therefore, it is preferable to use the endoscopes in AI diagnosis taking these facts into consideration.

References

Hirasawa T, Aoyama K, Tanimoto T, et al: Application of artificial intelligence using a convolutional neural network for detecting gastric cancer in endoscopic images. Gastric Cancer 21; 653-660: 2018
Ikenoyama Y, Hirasawa T, Ishioka M, et al: Detecting early gastric cancer; Comparison between the diagnostic ability of convolutional neural networks and endoscopists. Dig Endosc 33; 141-150: 2021
Miyaki R, Yoshida S, Tanaka S, et al: A computer system to be used with laser-based endoscopy for quantitative diagnosis of early gastric cancer. J Clin Gastroenterol 49; 108-115: 2015
Horiuchi Y, Aoyama K, Tokai Y, et al: Convolutional neural network for differentiating gastric cancer from gastritis using magnified endoscopy with narrow band imaging. Dig Dis Sci 65; 1355-1363: 2020
Ling T, Wu L, Fu Y, et al: A deep learning-based system for identifying differentiation status and delineating the margins of early gastric cancer in magnifying narrow-band imaging endoscopy. Endoscopy 53; 469-477: 2021
Hamada K, Kawahara Y, Tanimoto T, et al: Application of convolutional neural networks for evaluating the depth of invasion of early gastric cancer based on endoscopic images. J Gastroenterol Hepatol 37; 352-357: 2022

Introduction

Recently, artificial intelligence (AI)-based diagnostic support tools for early gastric cancer have been developed and reported to be superior to endoscopists (accuracy: 70% and sensitivity: 85%).¹ However, These tools could not guarantee reliable diagnostic accuracy in clinical practice. In previous reports, magnifying endoscopy has been recognized as superior to white-light imaging endoscopy for endoscopic diagnosis of early gastric cancers. Therefore, we developed an AI-based diagnostic support system using magnifying endoscopy with narrow band imaging (M-NBI) for early gastric cancers² and improved its application in clinical practice using nondegraded images obtained from selective depth inspection (SDI) output for real-time diagnosis. This study is the first to create an AI-based diagnostic support system using non-degraded M-NBI images obtained from SDI output, as several previous papers have analyzed damaged M-NBI images that were converted from various recording devices.^3,4

Aims & Methods

We built a new AI system (M-NBI AI) which was assembled using four different image recognition models with non-degraded M-NBI images (normal mucosa:1776 images, gastric cancers: 1947 images, and non-neoplastic lesions: 2451 images). We evaluated the performance of this M-NBI AI for diagnosis of early gastric cancer. The diagnostic performances of M-NBI AI and endoscopists (five specialists and five non-specialists) were compared using non-degraded M-NBI images (three images per lesion) of 50 early gastric cancer and 50 non-cancerous lesions. Early gastric cancers included histopathological types of well, moderately, and poorly differentiated adenocarcinomas and signet-ring cell adenocarcinomas. The primary outcome was the superiority of the M-NBI AI (based on accuracy) over the endoscopists. The secondary outcome was the superiority of the M-NBI AI (based on sensitivity) over the endoscopists.

Results

	M-NBI AI	Endoscopists
		Specialists	Non-specialists
Accuracy, % (95% CI)	87.0 (0.80–0.94)	80.0 (0.72–0.88)	71.0 (0.62–0.79)
Sensitivity, % (95% CI)	86.3 (0.77–0.96)	74.2 (0.63–0.85)	70.6 (0.58–0.83)
Specificity, % (95% CI)	88.0 (0.79–0.97)	89.0 (0.80–0.99)	71.0 (0.59–0.84)
PPV, % (95% CI)	88.0 (0.79–0.97)	92.0 (0.84–0.99)	72.0 (0.60–0.84)
NPV, % (95% CI)	86.0 (0.76–0.96)	68.0 (0.55–0.81)	70.0 (0.57–0.83)
PPV, positive predictive value; NPV, negative predictive value

The M-NBI AI was superior to the non-specialists based on accuracy (87.0%, 95%CI: 0.80–0.94 versus 71.0 %, 95%CI: 0.62–0.79; p<0.01), without being superior to the specialists (87.0%, 95%CI: 0.80–0.94 versus 80.0%, 95%CI: 0.72–0.88; p=0.18). The M-NBI AI was superior to the specialists and non-specialists based on sensitivity (86.3%, 95%CI: 0.77–0.96 versus 74.2%, 95%CI: 0.63–0.85 and 70.6%, 95%CI: 0.58–0.83; p<0.05, respectively).

Conclusion

The M-NBI AI was created as a new AI system that could be implemented for current endoscopy systems as a real-time diagnostic support tool for early gastric cancers. It demonstrated a robust performance and may indicate equalization of diagnostic accuracy and superadded effect over the white-light endoscopy for diagnosis of early gastric cancers in clinical practice.

References

1. Ishioka M et al. Dig Endosc. 2022. doi: 10.1111/den.14455.
2. Ueyama H et al. J Gastroenterol Hepatol. 2021;36(2):482-489.
3. Horiuchi Y et al. Dig. Dis. Sci. 2020;65:1355–1363.
4. He X et al. Gastrointest Endosc. 2022;95(4):671-678.

Introduction

There are several reports about the diagnosis of early gastric cancers (EGCs) using artificial intelligence (AI) systems, but satisfactory data for sensitivity and specificity are not yet available. Linked color imaging (LCI) produces high color contrast between malignant lesions and surrounding mucosa compared to white light imaging (WLI). Therefore, LCI combined with AI (LCI-AI) has the potential to further improve the diagnosis of gastric cancer compared with WLI combined with AI (WLI-AI). We developed three new LCI-AI models to determine their sensitivity and specificity for the diagnosis of EGCs.

Aims & Methods

Three models including a convolutional neural network (CNN) model (high resolution net (HRNet)), a transformer model (Segformer) and a hybrid (combined CNN and transformer) model (Uniformer), were based on MMSegmentation. When each model recognized a gastric cancer lesion from test image data, it was shown with the existing position as a blue area on WLI images and green area on LCI images. Non-neoplastic mucosa was shown as a black area on both images. Endoscopic images using WLI and LCI (LASEREO 7000 system series and scopes of EG-L600ZW and EG-L600ZW7, Fujifilm Corporation, Tokyo, Japan) were collected from patients with EGCs from January 2015 to December 2018 at Jichi Medical University Hospital. A total of 1970 images including 1179 WLI and 791 LCI images of 236 EGCs were used for training the three LCI-AI models. Another set of 414 EGC images was selected for validation from 43 cases. Primary and secondary endpoints were the sensitivity and specificity of each AI model, respectively.
The image-based performance metric was based on Dice loss, calculated by subtracting the mean Dice similarity score. A Dice score >20% was defined as the correct diagnosis of cancer because EGCs sometimes have unclear demarcation lines. The correct diagnosis of a non-malignant lesion was defined as no reactive black areas in surrounding mucosae. Each case had from four to 21 images. Therefore, the median Dice score was used as a representative value in each case to evaluate performance. All three models’ programming and data analysis were supported by AI Medical Service Inc. (Tokyo, Japan).

Results

Analysis of lesions in all EGC images showed both sensitivity (82.8%) and specificity (83.4%) with LCI-AI to be high and similar to those with WLI-AI (83.1% and 84.3%, respectively). Sensitivities of LCI-AI/ WLI-AI among HRNet, SegFormer and UniFormer were 80.9%/81.6% (p=0.625), 83.8%/83.8% (p=0.224) and 83.8%/83.8% (p=0.450), respectively and specificities were 80.0%/86.0% (p=0.120), 86.4%/84.8% (p=0.672) and 86.8%82.0%/ (p=0.214), respectively. Representative Dice values among cases with EGCs showed sensitivities for LCI-AI using HRNet, SegFormer and UniFormer to be extremely high, at 92.7% 90.9% and 93.0%, respectively, although not significantly different when compared to those using WLI-AI, 84.6% (p=0.67), 85.4% (p=0.15) and 80.5% (p=0.19), respectively. Higher sensitivities were found regardless of lesion size, morphology, and Helicobacter pylori status. All sensitivities using LCI-AI were higher than 80% whereas those using WLI-AI varied from 64% to 89%.

Conclusion

LCI-AI showed high sensitivity and specificity for the diagnosis of EGCs using all three models tested, suggesting potential for practical clinical use in the near future.

Disclosure

Hiroyuki Osawa and Hironori Yamamoto have consultant relationships with Fujifilm Corporation and have received honoraria, grants and royalties from the company. Hiroyuki Osawa has a consultant relationship with AI Medical survice inc. and has recieved honoraria.This study received no funding from Fujifilm Corporation and AI Medical survice inc. nor any other funding source.

Introduction

Although diagnosis of gastric intestinal metaplasia (IM) using image-enhanced endoscopy is spreading, it has the disadvantage that special training is required and the accuracy depends on individual ability and experience. Therefore, we developed and validated an artificial intelligence (AI) model for diagnosing IM, and subsequently, compared its diagnostic accuracy with that of endoscopists.

Aims & Methods

This study included 6,487 white-light endoscopic (WLE) images from 1,454 patients with histologically confirmed IM from a single tertiary referral center in Korea between 2016 and 2020 (training/validation/test set, n=5,764/650/91, respectively). Histopathological diagnoses, including the presence and grade of IM, were set as the gold standard. The performance parameters of the AI-assisted IM diagnostic (AI-IM) model were compared with those of visual diagnoses by independent endoscopists (11 experts and 20 non-experts).

Results

The AI-IM model showed an accuracy of 0.813 (95% confidence interval (CI), 0.718–0.887), sensitivity of 0.891 (95% CI, 0.788–0.955), specificity of 0.630 (95% CI, 0.424–0.806), area under the curve of 0.760 (95% CI, 0.659–0.843), and weighted average F1-score of 0.809 (95% CI, 0.680–0.897). On comparative analysis, the diagnostic accuracy of the AI-IM model was superior to both experts (0.711 [95% CI, 0.702–0.720], P=0.017) and non-experts (0.664 [95% CI, 0.654–0.673], P<0.001). Other parameters of performance were also found to be significantly better.

Conclusion

The AI-IM model shows favorable diagnostic accuracy compared to both experts and non-experts in diagnosing gastric IM based on WLE images.

Introduction

The depth of early gastric cancer is an important factor in determining the treatment plan such as endoscopic resection or surgery. We developed an artificial intelligence (AI) program using deep neural networks to diagnose mucosal and submucosal cancers using endoscopic images. Performance evaluation tests were conducted for clinical applications.

Aims & Methods

A validation dataset was established comprising 204 patients with early gastric cancer who underwent endoscopic resection or surgery at our hospital during 2018–2021. The lesion sites were the upper part in 16 (8%), middle part in 106 (52%), and lower part in 82 (40%) patients. The median lesion size (range) was 14 (2–90) mm; the lesion depth was mucosal cancer in 136 (67%) and submucosal cancer in 68 (33%) patients. Five white-light images per case at a minimum were used (median: 8, range: 5–22). The primary endpoint was sensitivity of mucosal cancer, defined as the lower limit of the 95% confidence interval (CI) > 75%, based on previous reports on the depth diagnosis of early gastric cancer. The secondary endpoints were specificity and accuracy.

Results

Sensitivity of this program for the diagnosis of mucosal cancer was 91.2% (95% CI: 85.2%–94.9%), indicating the usefulness of this program. Specificity was 66.2% (95% CI: 54.3%–76.3%), and accuracy was 82.8% (95% CI: 77.1%–87.4%).

Conclusion

The AI program developed for diagnosing the depth of early gastric cancer is useful and expected to be applied clinically in the near future.

Introduction

Gastric cancer has been identified as a serious global public health problem, and early detection of gastric cancer is an effective means of curbing the risk of gastric cancer. However, accurate, efficient and economical detection of early gastric cancer (EGC) remains a challenge.

Aims & Methods

We aim to fully explore the characteristics of various endoscopic images and EGC diagnostic requirements to boost the accuracy and efficiency of deep learning-based EGC detection and enable its public availability, so as to improve the survival rate and life quality of patients. A new lightweight EGC fusion diagnosis framework based on heterogeneous neural networks (FusionWN-Net) is proposed. FusionWN-Net consists of two stages: Model construction and Decision-level fusion. In the Model construction, two lightweight EGC classification models (WLI-Net and NBI-Net) are designed for the imaging characteristics of WLI and NBI to obtain accurate unimodal diagnostic results while guaranteeing low computational costs; in the Decision-level fusion, the Dempster-Shaffer algorithm is utilized to further fuse the diagnostic results of WLI-Net and NBI-Net to reach the final diagnosis. The proposed framework is trained and evaluated on a dataset of gastric endoscopic images collected and established from 1,055 patients of six hospitals, including Fuzhou University Affiliated Provincial Hospital. On this basis, we use Android Studio to develop a portable EGC diagnosis system (PEGCS) which can run smoothly on the mobile device.

Results

Extensive experiments show that the proposed framework exhibits superior performance in terms of both computational efficiency and diagnostic reliability. WLI-Net achieves 83.31% diagnostic accuracy with only 3.182M parameters and 0.471G FLOPs, and 37.04 FPS on the mobile device; NBI-Net achieves 82.85% diagnostic accuracy with 3.596M parameters and 0.609G FLOPs, and 24.39 FPS on the mobile device; the further obtained FusionWN-Net increases the accuracy to 86.52% with 6.778M parameters and 1.080G FLOPs, reaching the FPS of 14.08 on the mobile device, which is expected to be used in clinical practice.

Conclusion

We proposed WLI-Net and NBI-Net that can be used for white light endoscopic images and narrow band endoscopic images respectively, and a lightweight and efficient multimodal framework for EGC diagnosis, which are tailored for the unique characteristics and diagnostic requirements of different endoscopic images. Besides, we established the PEGCS to put the framework into practice. This work not only takes a big step forward in improving the EGC diagnostic accuracy, but also provides a general and efficient solution for obtaining high-quality EGC diagnosis results in resource-limited application scenarios, which demonstrates significant potential of clinical application.

Disclosure

The authors declare no conflicts of interest, financial or otherwise, related to the preparation or publication of this work.