Early Detection of Esophageal Cancer: Evaluating AI Algorithms with Multi-Institutional Narrowband and White-Light Imaging Data

doi:10.21203/rs.3.rs-4821229/v1

Download PDF

Article

Early Detection of Esophageal Cancer: Evaluating AI Algorithms with Multi-Institutional Narrowband and White-Light Imaging Data

https://doi.org/10.21203/rs.3.rs-4821229/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Esophageal cancer is one of the most common cancers worldwide, especially esophageal squamous cell carcinoma, which is often diagnosed at a late stage and has a poor prognosis. This study aimed to develop an algorithm to detect tumors in esophageal endoscopy images using innovative artificial intelligence (AI) techniques for early diagnosis and detection of esophageal cancer. We used white light and narrowband imaging data collected from Gachon University Gil Hospital, and applied YOLOv5 and RetinaNet detection models to detect lesions. We also verified the generalizability of the models using external data collected from various institutions. The results showed that the AI models showed high precision and sensitivity in white light and narrowband images, and the RetinaNet model performed well, analyzing not only polyps but also superficial esophageal cancer. This study demonstrates an effective method for detecting esophageal tumors through AI-based esophageal endoscopic image analysis and is expected to help researchers develop AI models that combine real-time image-based detection and synthesis algorithms to predict the depth of lesion invasion. These results and efforts are expected to significantly reduce misdiagnosis rates, enhance the effective diagnosis and treatment of esophageal cancer, and promote the standardization of medical services.

Health sciences/Gastroenterology

Health sciences/Medical research

Physical sciences/Engineering

esophageal cancer

endoscopy

deep learning

object detection

computer-aided diagnosis

Esophageal cancer is the eighth most common cancer worldwide and ranks among the top 10 fatal cancers [1]. However, almost all patients with esophageal adenocarcinoma are diagnosed at the end of the disease, and their prognosis is poor [2]. Currently, most diagnoses of esophageal squamous epithelial cell carcinoma are made using white-light imaging (WLI) endoscopy, and if dysplastic tissue is detected early, it can be treated with endoscopic mucosal resection and radio-frequency ablation [3]. Therefore, early detection and diagnosis are important for the survival and prognosis of patients with esophageal cancer [4]. Early diagnosis using WLI alone is difficult [5]. Instead of the iodine staining method, which induces problems such as chest pain, discomfort, and increased procedure time, a useful technique was used to identify the structure by emphasizing the microvessels on the surface of the esophageal squamous cell carcinoma using narrowband imaging (NBI) [6]. NBI can help detect and diagnose early esophageal squamous epithelial cell carcinoma [7]. The complexity of esophageal squamous epithelial cell carcinoma, characterized by its much more refined and intricate shape compared to polypoid lesions, makes its accurate detection all the more challenging. [8]. The use of conventional endoscopes is limited because they cannot easily discern the subtle changes in the initial lesion nor the number of biopsies, and lack high-definition imaging [9]. In addition, variability arises because of repeated and varied factors (experience, condition, fatigue, and mistakes) for the most important lesions [10]. Hence, the diagnostic accuracy of endoscopy may be reduced, and variations may occur depending on the diagnosis made by an inexperienced specialist [11, 12]. Thus, diagnostic assistance using artificial intelligence (AI) technology is needed to improve the quality of healthcare services and reduce the occurrence of medical errors by diagnosing and assisting medical staff in the medical field [13].

Recently, AI technology utilizing deep learning (DL) with Convolutional Neural Networks (CNNs) has been applied in the medical field for the detection of various lesions in endoscopic images [14]. It has shown excellent results in the diagnosis and detection of lesions in the stomach, small intestine, and colon. Diagnosis using AI can help medical staff detect lesions early [15, 16, 17]. Wang et al. developed a deep learning algorithm to evaluate the differences in polyp and adenoma detection performances through colonoscopy and verified its effectiveness [18]. Xu et al. developed a real-time classification and detection architecture for gastric polyps through gastroscopy that demonstrated excellent accuracy even in small polyp detection [19]. Goda et al. compared the predicted histological depth of esophageal squamous cell carcinoma by dividing it into mucosal or submucosal carcinoma between three endoscopic techniques and found that narrow-band imaging could reduce the risk of overestimation compared to other endoscopes [20]. Nakagawa et al. developed an AI system based on a single-shot multi-detector architecture for evaluating superficial squamous cell carcinoma and demonstrated performance similar to that of experienced endoscopists by diagnosing depth of invasion [21]. Although there have been many CNN-based studies on the detection and diagnosis of lesions in various organs, the medical data for esophageal squamous epithelial cell carcinoma is limited compared to other data, leading to issues such as overfitting and performance degradation on new lesion images [22]. Due to these technical difficulties, few studies have compared and evaluated white-light and narrowband images by applying deep learning for esophageal squamous epithelial cell carcinoma [23]. Therefore, the usefulness of applying detection methods needs to be evaluated and analyzed by collecting various esophageal squamous epithelial cell cancer data om white-light and narrowband images of AI-based multicenters.

In this study, we propose an AI algorithm that assists medical staff in the early detection and diagnosis of esophageal squamous epithelial cell carcinoma by analyzing esophageal endoscopy information based on collected data. In addition, to evaluate generalizability, the performance of the system was verified using multicenter data [24]. This system can be a useful tool for warning medical staff when dysplastic lesions are detected during esophageal endoscopy by overcoming the lack of generalizability compared with the results detected by endoscopy in one institution [25]. A deep learning algorithm for early detection of esophageal cancer, utilizing multicenter data from narrowband and white-light imaging, can be an excellent method with the potential to enhance the efficiency and accuracy of diagnosis and treatment.

Data Acquisition and Preprocessing

The data used in this study were obtained from 2,674 still images of 619 patients who underwent esophageal endoscopy (WLI) from January 2016 to June 2020 at Gachon University Gil Hospital and 480 still images of 121 patients who underwent esophageal endoscopy (NBI). This study received approval from the Gachon University Gil Hospital Clinical Research Ethics Review Committee, and the need for informed consent was waived due to the retrospective nature of the study (IRB No. GDIRB2020-316). All experimental protocols were performed in accordance with the relevant guidelines and regulations of the Declaration of Helsinki.

In the case of the WLI learning dataset, 2,674 sheets of normal, tumor-free, and tumor data from 619 patients were analyzed in a ratio of 8:1:1, divided into 1,925 sheets of learning data from 477 patients, 347 sheets of verification data from 60 patients, and 402 sheets of evaluation data from 82 patients, as shown in Table 1. In the case of the NBI learning dataset, based on the collected data, the dataset consisted of 480 sheets of 121 patients with tumors, divided into 374 sheets of learning data from 97 patients, 37 sheets of verification data from 12 patients, and 69 sheets of evaluation data from 12 patients, as shown in Table 1. As the WLI and NBI data had different horizontal and vertical ratios, all the images were resized to 640×640 pixels and used in the experiment.

Table 1

Number of images collected for WLI/NBI data
	White-light imaging			Narrowband imaging
	Normal	Cancer	Total	Cancer	Total
Train	238 (1,037)	239 (888)	477 (1,925)	97 (374)	97 (374)
Validation	30 (178)	30 (169)	60 (347)	12 (37)	12 (37)
Test	53 (201)	29 (201)	82 (402)	12 (69)	12 (69)

To learn and verify the deep learning model, the ground truth was obtained by labeling the location of the lesion. Using the ImageJ labeling software, a region of interest in the form of a rectangle, including the entire shape of the tumor, was displayed through a specialized inspection process. Among the collected data, one image was randomly selected for each image type and is presented in Fig. 1 with labels.

Configuration of Deep Learning Detection Models

The YOLOv5 model, which is a single-stage object detection framework, was applied to detect tumors in esophageal endoscopy images, as shown in Fig. 2 [26]. YOLO, which was used as a feature extraction model, predicts multiple areas in an image simultaneously using one convolution network and analyzes the class probability through a single regression. Its learning speed is faster because there is no complex pipeline in the model, and its detection performance is better than that of the R-CNN series model [27]. The parameters were normalized as the width and height of the image based on the ratio of the width and height of the bounding boxes. The prediction result of YOLO determines the final prediction label based on the prediction annotation coordinates and class probability.

Second, the RetinaNet model of the single-stage object detection framework, which introduced the concept of focus loss for the first time, was applied, as shown in Fig. 2 [28]. RetinaNet comprises a backbone network and two subnetworks that perform classification and bounding-box regression, respectively. The backbone, which is a publicly available open convolution network, calculates a convolution feature map for the entire area of the input image. The first subnet performed object classification using the convolution output of the backbone. The second subnet obtained the coordinates of the bounding box (offset between the anchor and the reference point) through convolution at the backbone output.

Experiment Setup

The experimental environment of this study used a system consisting of two NVIDIA GeForce GTX 2080 Ti (NVIDIA, Santa Clara, CA, USA) graphics processing units and an Intel® Xeon® Gold 6238 CPU @ 2.10 GHz and 32 GB of RAM and was executed on the Ubuntu 16.04 operating system. TensorFlow (version 1.14.0), PyTorch (version 1.7.1), Keras (version 2.2.4), and Python (version 3.6.12) were used for the deep learning.

Deep Learning Model Parameters and Evaluation Index

In this study, YOLOv5l was used among five model sizes, from YOLOv5sl to YOLOv5xl. The prediction and learning conditions of YOLOv5 were set to 200 epochs and a batch size of 16 using an image size optimization algorithm of 640×640 and a learning rate of 1e-3 (Adam). An early stopping algorithm was applied to prevent overfitting. The prediction and learning conditions of RetinaNet were set to 200 epochs and a batch size of 1 using an image size (learning rate) optimization algorithm of 640×640 and a learning rate of 1e-5. An early stopping algorithm was applied to prevent overfitting.

The learned detection model was compared and analyzed using performance evaluation indicators such as precision, sensitivity, and false positives per image (FPPI). The confidence score is an index that can determine the class classification and position detection results of the detected boundary box and is obtained by multiplying the probability of the class predicted by the model to be correct for the object detected with the intersection over union (IoU) value. A true positive (TP) is the case of correctly detecting the tumor location obtained through the tumor location detection model; a false positive (FP) is the case of detecting the location without the tumor; and a false negative (FN) is the case of failure to detect the tumor. Using the confusion matrices, several performance indicators were calculated as (1), (2), (3), and (4). Precision is the ratio of what the model correctly identifies among the predictions of a lesion, and sensitivity is the ratio of what the model predicts among the data with actual lesions. The ratio of the number of FP images detected per image was used for the FPPI, which indicates that the scale fluctuation was very wide, depending on the data. To compare and evaluate the detection performance, the change according to the adjustment of the parameter through the precision–recall curve with an exchange relationship was graphically represented.

$$\:Precision\:=\:\frac{TP}{TP+FP}$$

$$\:Sensitivity\:=\:\frac{TP}{TP+FN}\:$$

$$\:FPPI\:=\:\frac{FP}{Number\:of\:images}$$

$$\:IoU=\frac{Avea\:of\:Overlap}{Area\:of\:Union}$$

Additional Experiments

To test the generalizability of the model, external validation was performed using images acquired from patients who underwent esophageal endoscopy (WLI and NBI) at Kyunghee Medical Center, Korea University Anam Hospital, and Hallym University Sacred Heart Hospital. For the WLI dataset, the detection performance of the model was tested using data from 112 tumors from Kyunghee Medical Center, 353 tumors from Korea University Anam Hospital, and 23 tumors from Hallym University Sacred Heart Hospital. For the NBI learning dataset, the detection performance of the model was tested using data from 29 tumors from Kyunghee University Hospital, 192 tumors from Korea University Anam Hospital, and 13 tumors from Hallym University Sacred Heart Hospital.

Performance Comparison of Detection Network Models

Based on the presence or absence of lesions in the esophageal endoscopy image, two classes of data—normal and with lesions—were designated, and the results detected by the AI model were analyzed. Esophageal lesions were defined as true, and no esophageal lesions were defined as false. When the IoU value between the prediction and correct answer areas was 0.5, the prediction was considered successful.

To confirm the precision, sensitivity, and FPPI according to the compliance threshold, the performances of the esophageal cancer detection models YOLOv5 and RetinaNet were compared and analyzed for the detection results with a threshold value of 0.1 or more, as shown in Table 2. In the WLI dataset, the YOLOv5 model detected images with a precision of 93.7%, a sensitivity of 89.9%, and an FPPI of 6%. The RetinaNet model detected images with a precision of 96.1%, a sensitivity of 88.4%, and an FPPI of 3.5%. In the NBI dataset, the YOLOv5 model detected images with a precision of 86.5%, a sensitivity of 84.0%, and an FPPI of 13%. The RetinaNet model detected images with a precision of 98.4%, a sensitivity of 91.3%, and an FPPI of 1.4%. From the WLI dataset, 402 evaluation data points were obtained, composed of 201 normal data points without tumors and 201 data points with tumors, and the performance of the detection model was evaluated. In the YOLOv5 model, 179 of the 201 data points with tumors were determined as data with tumors (TP), and 20 were determined as data without tumors (FN). Moreover, 12 of the 201 normal data points without tumors were determined to have tumors (FP). In the RetinaNet model, 176 of the 201 data points with tumors were determined as data with tumors (TP), and 23 were determined as data without tumors (FN). Moreover, 7 of the 201 normal data points without tumors were determined as data with tumors (FP). By showing an example of image detection in Fig. 3, the true detection results of the tumor location predicted by the detection model and the actual tumor location can be confirmed. In the NBI dataset, 69 evaluation data points with tumors were constructed to evaluate the performance of the detection model. In the YOLOv5 model, 58 of the 69 data points with tumors were determined as data with tumors (TP), and 11 were determined as data without tumors (FN). Nine normal data points without tumors were identified as data points with tumors (FP). In the RetinaNet model, 63 of the 69 data points with tumors were determined as data with tumors (TP), and six were determined as data without tumors (FN). A normal datum without tumors was determined to be a datum with tumors (FP). As shown in Fig. 4, the detection of FP and FN results for the tumor location predicted by the detection model and the actual tumor location can be confirmed from the internal data. FP results were obtained because of the prediction of shadows from normal data as lesions, which accounted for almost all cases. In addition, as shown in Fig. 4b, when the lesion was very small and far away, the nearby crystal area was predicted to be an FP. The main cause of the FN predicted results was esophageal inflammation in the mucous membrane, as shown in Fig. 4c. In addition, even when the lesion occupied the entire area, as shown in Fig. 4d, it could not be predicted. Figure 5 shows the overall performance of the model with a precision–recall curve for the internal data. In general, the closer the curve is to the upper-right corner, the better the performance of the model. The two detection models identify the positive classes well and simultaneously consider the number of negative classes incorrectly classified as positive. In the detection model, recall with a low FN ratio was more important.

Table 2

Performance evaluation metrics for detection models based on confidence thresholds from internal data
	Model	Precision (95% CI)	Sensitivity (95% CI)	FPPI (95% CI)
White-light imaging	YOLOv5	0.937 (0.892–0.951)	0.899 (0.857–0.924)	0.06 (0.043–0.075)
White-light imaging	RetinaNet	0.961 (0.88–0.984)	0.884 (0.80-0.954)	0.035 (0.016–0.043)
Narrowband imaging	YOLOv5	0.865 (0.824–0.913)	0.840 (0.763–0.88)	0.13 (0.07–0.267)
Narrowband imaging	RetinaNet	0.984 (0.951–0.99)	0.913 (0.842–0.944)	0.014 (0.008–0.035)

Through external verification, according to the presence or absence of lesions in the esophageal endoscopic image, two classes of data were designated as normal without lesions and with lesions, and the results detected by the AI model were analyzed. To confirm the precision, sensitivity, and FPPI according to the compliance threshold, the performances of the esophageal cancer detection models YOLOv5 and RetinaNet were compared and analyzed for the detection results with a threshold value of 0.1 or more, as shown in Table 3. In the WLI dataset, the YOLOv5 model detected images with a precision of 83.4%, a sensitivity of 79.4%, and an FPPI of 15.8%. The RetinaNet model detected images with a precision of 88.3%, a sensitivity of 70.2%, and an FPPI of 9.2%. In the NBI dataset, the YOLOv5 model detected images with a precision of 85.6%, a sensitivity of 71.3%, and an FPPI of 11.9%. The RetinaNet model detected images with a precision of 88.3%, a sensitivity of 81.1%, and an FPPI of 10.6%. In the WLI dataset, 488 evaluation data points with tumors were constructed to evaluate the performance of the detection model. In the YOLOv5 model, 387 of the 488 tumors were identified as data with tumors (TP) and 100 as data without tumors (FN). Moreover, 77 normal data points without tumors were identified as data with tumors (FP). In the RetinaNet model, 342 of the 488 tumors were determined as data with tumors (TP), and 145 were determined as data without tumors (FN). Moreover, 45 normal data points without tumors were identified as data with tumors (FP). By showing an example of image detection in Fig. 6, the true detection results of the tumor location predicted by the detection model and the actual tumor location can be confirmed.

Table 3

Performance evaluation metrics for detection models based on confidence thresholds from external data
	Model	Precision (95% CI)	Sensitivity (95% CI)	FPPI (95% CI)
White-light imaging	YOLOv5	0.834 (0.761–0.88)	0.794 (0.756–0.85)	0.158 (0.086–0.194)
White-light imaging	RetinaNet	0.883 (0.82–0.956)	0.702 (0.657–0.782)	0.092 (0.057–0.18)
Narrowband imaging	YOLOv5	0.856 (0.817–0.893)	0.713 (0.68–0.76)	0.119 (0.007–0.142)
Narrowband imaging	RetinaNet	0.883 (0.853–0.947)	0.811 (0.766–0.89)	0.106 (0.08–0.181)

In the NBI dataset, 288 evaluation data points with tumors were constructed to evaluate the performance of the detection model. In the YOLOv5 model, 167 of the 288 tumors were identified as data with tumors (TP), and 67 were determined as data without tumors (FN). Moreover, 28 normal data points without tumors were identified as data points with tumors (FP). In the RetinaNet model, 190 of the 288 tumors were determined as data with tumors (TP), and 44 were determined as data without tumors (FN). Moreover, 25 normal data points without tumors were identified as data points with tumors (FP). As shown in Fig. 7, the detection of FP and FN results for the tumor location predicted by the detection model and the actual tumor location can be confirmed from external data. FP results were obtained because of the prediction of shadows from normal data as lesions, which accounted for almost all cases. In addition, as shown in Fig. 7b, the overall lesion was predicted as an FP. The main cause of the result predicted as an FN was the presence of only a part of the lesion, as shown in Fig. 7c. The case of esophageal inflammation of the mucous membrane shown in Fig. 7d could not be predicted to be a lesion. Figure 8 shows the overall performance of the model with a precision–recall curve for external data. In general, the closer the curve is to the upper-right corner, the better the performance of the model. The two detection models identify the positive classes well and simultaneously consider the number of negative classes incorrectly classified as positive. In the detection model, recall with a low FN rate is more important. High precision can lead to low recall, indicating that the model misses most of the tumor data.

In this study, an AI algorithm was proposed to detect the location of esophageal tumors using multicenter data from esophageal endoscopic WLI and NBI tests. Normal data without areas of interest were also learned so that the AI learning model could reduce FPs and increase sensitivity in normal data. To increase sensitivity, which is particularly important in detection research, it is necessary to find as many objects as possible in the data. Some images have no special elements; hence, the FPPI is 0. However, other images have several objects that can be recognized as objects; hence, the FPPI can have a large value. The study was conducted by carefully analyzing the performance indicators. The confidence threshold was set to 0.25, and the IoU threshold was set to 0.5, to prevent the removal of additional bounding boxes. The detection model showed high precision and sensitivity, and normal and tumor data were classified with high accuracy. In addition, data were collected from various institutions to verify the relatively high generalization performance. The established database can be used as important data for CAD research and algorithm development for future endoscopies. Most of the related cases mentioned were mainly polyps to detect and diagnose lesions in various organs; however, in this study, lesion detection was performed in white-light and narrow-band images by analyzing not only polyps but also superficial esophageal cancer [14, 15, 16, 17, 18, 19, 20, 21].

To improve the performance of the model in the future, relearning through additional data collection and cross-verification processes is required to improve its reliability. In addition, the performance needs to be optimized by fine-tuning the parameters of the AI algorithm based on feedback from the verification [29]. It is determined that the area of the tumor location will stand out owing to the mitigation or overcoming of the limitations of the existing pre-processing, which will be helpful in learning performance. To detect morphological tumors, the performance of the tumor detection model can be improved by increasing its clinical suitability by performing post-treatment separately [30].

This study demonstrated the possibility of an effective method for detecting the location of esophageal tumors in AI-based esophageal endoscopy images. This can be advantageously used in the current endoscopy environment in the future by applying division and processing in frames based on real-time videos. Based on these studies, algorithmic weight reduction and optimization technologies can be implemented to enable real-time processing by improving the processing speed of each algorithm. In addition, it is thought that incorporating synthetic algorithm technology that generates data, such as a generative adversarial network, from narrowband images can help develop an AI model that predicts the depth of invasion of the lesion. These efforts are expected to contribute to the effective diagnosis and treatment of esophageal cancer and promote the standardization of medical services. This study demonstrated the possibility of effective research on tumor detection through deep-learning-based esophageal endoscopy. These results are expected to improve the quality of medical services, enable a more precise and fast diagnosis, and contribute to reducing the misdiagnosis rate by providing diagnostic support to medical staff.

Author Contribution

Y.S developed the algorithm, trained, and evaluated the models used in this study. J.W drafted the manuscript. H.N and J.W collected the dataset. K.G and Y.J participated in the study design and revised the manuscript. All the authors read and approved the final manuscript.

Acknowledgement

This work was supported by a grant from the Korean Gastrointestinal Endoscopy Research Foundation (2021 Investigation Grant), and by the Gachon University Gil Medical Center (Grant number: FRD2022-12), and by the Gachon University research fund of 2023(GCU-202308020001)

Data Availability

The esophageal endoscopic image data used to support the findings of this study are available upon request from the corresponding author.

Pohl, H., Sirovich, B & Welch HG. Esophageal adenocarcinoma incidence: are we reaching the peak? Cancer epidemiology, biomarkers & prevention. 19 (2010) 1468–1470. https://doi.org/10.1158/1055-9965.EPI-10-0012.
Hur, C., et al. Trends in esophageal adenocarcinoma incidence and mortality. Cancer. 119 (2013) 1149–1158. https://doi.org/10.1002/cncr.27834.
Behrens, A., et al. Barrett’s adenocarcinoma of the esophagus: better outcomes through new methods of diagnosis and treatment. Deutsches Ärzteblatt International, 108 (2011) 313. https://doi.org/10.3238%2Farztebl.2011.0313.
Bird-Lieberman, E.L. & Fitzgerald, RC. Early diagnosis of oesophageal cancer. British journal of cancer, 101 (2009) 1–6. https://doi.org/10.1038/sj.bjc.6605126.
Nagami, Y., et al. Usefulness of non-magnifying narrow-band imaging in screening of early esophageal squamous cell carcinoma: a prospective comparative study using propensity score matching. The American journal of gastroenterology, 109 (2014) 845. https://doi.org/10.1038/ajg.2014.94.
Lee, YC., et al. Transnasal endoscopy with narrow-band imaging and Lugol staining to screen patients with head and neck cancer whose condition limits oral intubation with standard endoscope (with video). Gastrointestinal endoscopy, 69 (2009) 408–417. https://doi.org/10.1016/j.gie.2008.05.033.
Kuraoka, K., et al. Early esophageal cancer can be detected by screening endoscopy assisted with narrow-band imaging (NBI). Hepato-gastroenterology, 56 (2009) 63–66. https://doi.org/10.11404/jsgcs.46.551.
Li, H., et al. Single-Image-Based Deep Learning for Segmentation of Early Esophageal Cancer Lesions. (2023) arXiv preprint arXiv:2306.05912. https://doi.org/10.1109/TIP.2024.3379902.
Pennathur, A., Gibson MK., Jobe, BA., Luketich JD. Oesophageal carcinoma. The Lancet, 381 (2013) 400–412. https://doi.org/10.1016/S0140-6736(12)60643-6.
Emery, JD., et al. The role of primary care in early detection and follow-up of cancer. Nature reviews Clinical oncology, 11 (2014) 38–48. https://doi.org/10.1038/nrclinonc.2013.212.
Forbes, LJ., Warburton F., Richards MA., Ramirez AJ. Risk factors for delay in symptomatic presentation: a survey of cancer patients. British journal of cancer, 111 (2014) 581–588. https://doi.org/10.1038/bjc.2014.304.
Walter, FM., et al. Symptoms and other factors associated with time to diagnosis and stage of lung cancer: a prospective cohort study. British journal of cancer, 112 (2015) S6-S13. https://doi.org/10.1038/bjc.2015.30.
Kudo, SE., et al. Artificial intelligence-assisted system improves endoscopic identification of colorectal neoplasms. Clinical Gastroenterology and Hepatology, 18 (2020) 1874–1881. https://doi.org/10.1016/j.cgh.2019.09.009.
Horie, Y., et al. Diagnostic outcomes of esophageal cancer by artificial intelligence using convolutional neural networks. Gastrointestinal endoscopy, 89 (2019) 25–32. https://doi.org/10.1016/j.gie.2018.07.037.
Ikenoyama, Y., et al. Detecting early gastric cancer: Comparison between the diagnostic ability of convolutional neural networks and endoscopists. Digestive Endoscopy, 33 (2021) 141–150. https://doi.org/10.1111/den.13688.
Ribeiro, E., UhI, A., Wimmer, G., Hafner, W. Exploring deep learning and transfer learning for colonic polyp classification. Computational and mathematical methods in medicine(2016). https://doi.org/10.1155/2016/6584725.
Ding, Z., et al. Gastroenterologist-level identification of small-bowel diseases and normal variants by capsule endoscopy using a deep-learning model. Gastroenterology, 157 (2019) 1044–1054. https://doi.org/10.1053/j.gastro.2019.06.025.
Wang, P., et al. Development and validation of a deep-learning algorithm for the detection of polyps during colonoscopy. Nature biomedical engineering, 2 (2018) 741–748. https://doi.org/10.1038/s41551-018-0301-3.
Zhang, X., et al. Real-time gastric polyp detection using convolutional neural networks. PloS one, 14 (2019) e0214133. https://doi.org/10.1371/journal.pone.0214133.
Goda, K., et al. Magnifying endoscopy with narrow band imaging for predicting the invasion depth of superficial esophageal squamous cell carcinoma. Diseases of the Esophagus, 22 (2019) 453–460. https://doi.org/10.1111/j.1442-2050.2009.00942.x.
Nakagawa, K., et al. Classification for invasion depth of esophageal squamous cell carcinoma using a deep neural network compared with experienced endoscopists. Gastrointestinal endoscopy, 90 (2019) 407–414. https://doi.org/10.1016/j.gie.2019.04.245.
Malick, A., Soroush, A & JA. Abrams. Esophageal dysbiosis and esophageal squamous cell carcinoma. Esophageal Disease and the Role of the Microbiome (2023) 91–114. https://doi.org/10.1016/B978-0-323-95070-1.00014-X.
Wang, ZX., et al. Linked color imaging vs Lugol chromoendoscopy for esophageal squamous cell cancer and precancerous lesion screening: A noninferiority study. World Journal of Gastroenterology, 29 (2023) 1899. https://doi.org/10.3748%2Fwjg.v29.i12.1899.
Ali, S., et al. Deep learning for detection and segmentation of artefact and disease instances in gastrointestinal endoscopy. Medical image analysis, 70 (2021) 102002. https://doi.org/10.1016/j.media.2021.102002.
Vallböhmer, D., et al. A multicenter study of survival after neoadjuvant radiotherapy/chemotherapy and esophagectomy for ypT0N0M0R0 esophageal cancer. Annals of surgery, 252 (2010) 744–749. https://doi.org/10.1097/SLA.0b013e3181fb8dde.
ultralytics/yolov5: v5. 0-YOLOv5-P6 1280 models, AWS, Supervise. ly and YouTube integrations. Zenodo (2021). https://ui.adsabs.harvard.edu/link_gateway/2021zndo… J/doi:10.5281/zenodo.4679653,
Jabir, B., Falih, N & Rahmani, K. Accuracy and efficiency comparison of object detection open-source models. International Journal of Online & Biomedical Engineering, 17 (2021). https://doi.org/10.3991/IJOE.V17I05.21833,
Lin, TY., Goyal, P., Girshick, R., He, K & Dollar, P. Focal loss for dense object detection. Proceedings of the IEEE international conference on computer vision, (2017) 2980–2988. https://doi.org/10.1109/iccv.2017.324.
Wang, CC., et al. A deep learning model for classification of endoscopic gastroesophageal reflux disease. International journal of environmental research and public health, 18 (2021) 2428. https://doi.org/10.3390/ijerph18052428.
Ali, S., et al. A deep learning framework for quality assessment and restoration in video endoscopy. Medical image analysis, 68 (2021) 101900. https://doi.org/10.1016/j.media.2020.101900.

No competing interests reported.

Download PDF

Reviews received at journal
26 Aug, 2024
Reviewers agreed at journal
22 Aug, 2024
Reviewers agreed at journal
14 Aug, 2024
Reviewers invited by journal
07 Aug, 2024
Editor assigned by journal
30 Jul, 2024
Editor invited by journal
30 Jul, 2024
Submission checks completed at journal
30 Jul, 2024
First submitted to journal
29 Jul, 2024

You are reading this latest preprint version

Early Detection of Esophageal Cancer: Evaluating AI Algorithms with Multi-Institutional Narrowband and White-Light Imaging Data

Status:

Version 1

Abstract

Figures

Introduction

Materials and Methods

Data Acquisition and Preprocessing

Configuration of Deep Learning Detection Models

Experiment Setup

Deep Learning Model Parameters and Evaluation Index

Additional Experiments

Results

Performance Comparison of Detection Network Models

Discussion

Declarations

Author Contribution

Acknowledgement

Data Availability

References

Additional Declarations

Status:

Version 1