Methods for using artificial intelligence models to assist in the detection of Barrett's esophageal mucosal lesions

Abstract

This invention provides a method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model. The method involves a processing unit of a computing device executing the following actions according to program instructions: analyzing an esophageal image under test using a Barrett's mucosal detection model to detect whether lesions are present in the image; and if lesions are detected, further locating and marking the position of the lesions. In other words, the method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model disclosed in this invention can automatically detect and mark areas in the esophageal image under test that are at risk of Barrett's esophageal mucosal lesions, thereby achieving the effects of precision medicine, rapid detection, and assisted interpretation of medical images.

Description

[Technical Field]

[0001] This invention relates to a method for predicting the risk of precancerous lesions, and more particularly to a method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model. [Previous Technology]

[0002] Note: Barrett's esophagus is a condition caused by long-term gastroesophageal reflux, resulting in the transformation of esophageal squamous epithelial cells into small intestinal columnar epithelial cells, clinically known as "squamous metaplasia into columnar epithelial cells." Studies indicate that patients with Barrett's esophagus have a 30-60 times higher risk of developing esophageal adenocarcinoma than the general population; therefore, Barrett's esophagus is considered a precancerous lesion.

[0003] Although Barrett's esophagus is considered to be associated with esophageal adenocarcinoma, not all Barrett's esophagus has a high risk of cancerous transformation. Studies indicate that only highly poorly differentiated Barrett's esophagus has the potential to transform into esophageal adenocarcinoma within 3 years. Since endoscopy cannot determine whether Barrett's esophagus is poorly differentiated, the current clinical practice is to perform pathological sections to confirm the differentiation status after the presence of symptoms of Barrett's esophagus. Furthermore, previous literature shows that the sectioning rate and diagnostic rate of Barrett's esophagus in Asian regions are poor, thus necessitating the development of artificial intelligence-assisted methods.

[0004] In addition, although there are studies that have revealed the technology of identifying early Barrett's esophagus tumors from esophageal images through machine learning, the previous technology was used to classify and identify esophageal tissues that have already developed tumors. It was mainly used to identify and detect the location of dysplasia in Barrett's esophagus. Therefore, the conventional technology has the shortcoming that it cannot detect when there is no dysplasia in Barrett's esophagus. [Summary of the Invention]

[0005] The main objective of this invention is to provide a method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model. This method can automatically detect the image to be tested at the gastroesophageal junction and locate and mark the location of the lesion through a Barrett's mucosal detection model, so as to assist medical personnel in determining whether a biopsy or further treatment should be performed, thereby achieving the effects of precision medicine and early prevention.

[0006] Another objective of this invention is to provide a method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model, which can be used as a tool for telemedicine, thereby improving the quality of medical care in remote areas and reducing medical costs.

[0007] Therefore, in order to achieve the above objectives, the method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model disclosed in this invention analyzes an esophageal image to be tested using a Barrett's mucosal detection model to detect and locate areas in the esophageal image that have Barrett's esophageal mucosal lesions or are at high risk, so as to produce a prediction result. That is, when the Barrett's mucosal detection model detects areas in the esophageal image to be tested that have Barrett's esophageal mucosal lesions at risk, the prediction result includes a prediction box to select the areas in the esophageal image to be tested that have been identified and detected to have Barrett's esophageal mucosal lesions, and can display the probability that the areas selected by the prediction box have Barrett's esophageal mucosal lesions or their characteristics.

[0008] The esophageal image to be tested is an endoscopic image located at the junction of the stomach and esophagus.

[0009] In one embodiment of the present invention, the Barrett's mucosal detection model is obtained by the following steps:

[0010] Step A-1: ​​Obtain a Barrett's esophageal mucosa dataset and a non-Barrett's esophageal mucosa dataset, wherein the Barrett's esophageal mucosa dataset contains multiple esophageal image data of individuals determined to have Barrett's esophageal mucosa; and the non-Barrett's esophageal mucosa dataset contains esophageal image data of individuals without Barrett's esophageal mucosa.

[0011] Step A-2: Take at least one set of Barrett's esophageal mucosa data and at least one set of non-Barrett's esophageal mucosa data as a training data set, input the training data set into an integrated object detection and classification model for calculation and analysis, and produce the Barrett's mucosa detection model.

[0012] The esophageal imaging data includes an endoscopic image, a dynamic video, a histological image and / or a static image.

[0013] The esophageal imaging data with Barrett's esophageal mucosa includes a lesion marker, such as a selection box.

[0014] In another embodiment of the present invention, step A-2 further includes taking another set of Barrett's esophageal mucosa data and another set of non-Barrett's esophageal mucosa data as a test data set, and inputting the test data set into the Barrett's mucosa detection model for predictive analysis, so as to confirm the prediction accuracy of the Barrett's mucosa detection model.

[0015] In another embodiment of the present invention, the method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model further includes a step D, wherein after step C, the prediction result is included in the Barrett's esophageal mucosal dataset or the non-Barrett's esophageal mucosal dataset.

Implementation Method

[0016] The present invention provides a method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model. The method involves a processing unit of a computing device executing the following actions according to program instructions: analyzing an esophageal image to be tested using a Barrett's mucosal detection model to determine whether there is tissue suspected to be Barrett's esophageal mucosa in the esophageal image. If the Barrett's mucosal detection model detects tissue suspected to be Barrett's esophageal mucosa in the esophageal image, the Barrett's mucosal detection model will identify the area as having a risk of Barrett's esophageal mucosa and locate and mark the area to achieve the effect of rapid detection and auxiliary interpretation of medical images.

[0017] Furthermore, the Barrett's mucosa detection model is obtained by performing calculations and analysis on a training dataset using an integrated object detection and classification model. The training dataset includes multiple esophageal images of Barrett's esophageal mucosa confirmed by pathological sections, and the non-Barrett's esophageal mucosa dataset includes esophageal images of non-Barrett's esophageal mucosa.

[0018] The term "Barrett's esophageal mucosa" or "Barrett's esophagus" refers to the change in cell morphology of the esophageal mucosa near the cardia of the stomach, from squamous epithelial cells (flat cells) to columnar cells of the small intestine.

[0019] The term “lesion location” or “lesion site” in this invention refers to the site in esophageal imaging that is at risk of Barrett’s esophageal mucosa.

[0020] The term “suspected Barrett’s esophageal mucosa” refers to esophageal tissue that has Barrett’s esophageal mucosa or its symptoms, and requires further pathological confirmation to determine whether it has become diseased.

[0021] The term "esophageal image" refers to an image showing the esophageal tissue at the gastroesophageal junction, wherein the image may be an endoscopic image, a video, a histological image or a static image.

[0022] The term "Ensemble Object Detection and Classification Model" refers to an algorithm that combines at least two object detection and / or classification models.

[0023] The term "computing device" refers to a device used to execute programs, perform data processing and / or control hardware, such as servers, computers, cloud platforms, etc.

[0024] The term "processing unit" refers to the unit used to execute program instructions and perform calculations, also known as a processor.

[0025] The term "program instruction" refers to an executable logical step, such as model calculation or data retrieval. Generally, program instructions are stored in memory units, such as memory, and can be read and executed by processing units.

[0026] The term "object detection (OD)" refers to the detection of a predetermined location, such as a lesion site, in an esophageal image.

[0027] The term "Ensemble Method" or "Object Detection Ensemble Method" refers to fusing or combining the prediction results obtained from multiple object detection models to obtain a final prediction result with better prediction performance. Common object detection ensemble methods include Non-Maximum Suppression Ensemble (NMS Ensemble) and Weighted Box Fusion (WBF).

[0028] The term "lesion marker" refers to a mark used to mark the Barrett's esophageal mucosa or the area containing Barrett's esophageal mucosa in esophageal images of patients who have been diagnosed with Barrett's esophageal mucosa, such as lines, arrows, squares, circles, etc.

[0029] The following describes the technical features of the present invention and the effects it can achieve. Several embodiments are given below in conjunction with the accompanying drawings for detailed explanation.

[0030] The method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model disclosed in the first embodiment of the present invention comprises steps executed by the processing unit of a computing device according to program instructions. Specifically, the method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model includes the following steps:

[0031] Step A: Obtain a Barrett's mucosa detection model, as shown in Figures 1 to 3. This Barrett's mucosa detection model is established based on the following steps:

[0032] Step A-1: ​​Obtain a Barrett's esophageal mucosa dataset and a non-Barrett's esophageal mucosa dataset. The Barrett's esophageal mucosa dataset includes multiple esophageal imaging data classified as having Barrett's esophageal mucosa, such as esophageal imaging data determined by pathological sections or endoscopy. The non-Barrett's esophageal mucosa dataset includes esophageal imaging data of non-Barrett's esophageal mucosa, obtained from the normal esophagus of a patient with non-Barrett's esophageal mucosa. The esophageal imaging data includes dynamic and static images, such as static and dynamic images captured during endoscopic examination.

[0033] Furthermore, each of the esophageal images in the Barrett's esophageal mucosa dataset undergoes an image preprocessing procedure, that is, lesion marking is performed on the esophageal sites with Barrett's esophageal mucosa on each esophageal image. The lesion marking may be a circle, a box, or any mark that can highlight or distinguish the lesion site.

[0034] Step A-2: A portion of the Barrett's esophageal mucosa dataset and a portion of the non-Barrett's esophageal mucosa dataset are used as a training dataset, and the remaining portion of the Barrett's esophageal mucosa dataset and the remaining portion of the non-Barrett's esophageal mucosa dataset are used as a test dataset. The training dataset is input into an integrated object detection and classification model for computational analysis to produce the Barrett's mucosa detection model; and the test dataset is input into the Barrett's mucosa detection model to confirm the prediction accuracy of the Barrett's mucosa detection model.

[0035] Generally, the amount of data used as training data is greater than the amount of data used as testing data. For example, 90% of the data set of Barrett's esophageal mucosa and 90% of the data set of non-Barrett's esophageal mucosa can be used as training data, and the remaining 10% of each data set can be used as training data; or 80% of the data set of Barrett's esophageal mucosa and 80% of the data set of non-Barrett's esophageal mucosa can be used as training data, and the remaining 20% ​​of each data set can be used as training data; or 70% of the data set of Barrett's esophageal mucosa and 70% of the data set of non-Barrett's esophageal mucosa can be used as training data, and the remaining 30% of each data set can be used as training data.

[0036] In this embodiment, the Barrett's esophageal mucosa data set and the non-Barrett's esophageal mucosa data set are from the endoscopic imaging system of Chung Shan Medical University Hospital. The subjects include 64 patients suspected of having Barrett's esophageal mucosa by endoscopy, 91 patients confirmed by pathological sections to have Barrett's esophageal mucosa, and 31 patients with non-Barrett's esophageal mucosa. The images were collected from January 2023 to March 2025.

[0037] Further, as shown in Figures 2 and 3, the integrated object detection and classification model first analyzes the training dataset or the test dataset using multiple object detection modules to obtain an initial prediction result. The multiple object detection modules include YOLOv4, YOLOv5, YOLOv6, YOLOv7, YOLOv8, YOLOv9, YOLOv10, YOLOv11, and YOLOv12. The initial prediction result includes a class, a bounding box, and a confidence score. The class refers to the model's determination that the selected object belongs to the lesion location. The bounding box refers to the position and / or size of the object detected by the model, which is usually composed of four values. Then, the initial prediction results are combined using an object detection ensemble method. The confidence score refers to the model's confidence in this prediction result. Usually, when the confidence score is lower than the set confidence threshold, the prediction result is discarded. Then, the initial prediction results are integrated using Non-Maximum Suppression Ensemble (NMS Ensemble), and then evaluated and compared to produce the final prediction results. Commonly used evaluation metrics include recall, precision, and IoU (Intersection over Union). The comparison referred to here is the comparison of the detection performance of the results obtained by NMS Ensemble with other standard models, and the comparison metrics are recall, precision, mAP (mean Average Precision), etc.

[0038] In addition, as shown in Figure 2, the Barrett mucosal detection model can be further validated externally to achieve the purpose of validation and model optimization. The dataset used for external validation is an external validation dataset, and the data sources contained therein are different from the training dataset or the test dataset, such as other medical centers or medical testing units.

[0039] Step B: Obtain an esophageal image to be tested, wherein the esophageal image to be tested is an endoscopic image, which may be a continuous image or a segmented photograph extracted from the dynamic image.

[0040] Step C: Input the esophageal image to be tested into the Barrett's mucosa detection model. The Barrett's mucosa detection model analyzes the esophageal image to detect and locate areas in the esophageal image that have a risk of Barrett's esophageal mucosa or are suspected of having Barrett's esophageal mucosa, and produces a detection result. When the Barrett's mucosa detection model detects areas in the esophageal image to be tested that have a risk of Barrett's esophageal mucosa, the detection result includes a prediction box to detect and select areas in the esophageal image to be tested that are suspected of having Barrett's esophageal mucosa, and the area selected by the prediction box shows the probability that it has Barrett's esophageal mucosa or its characteristics. For example, the display method can be in the form of percentages, such as 92.5%, 50.5%, etc.; or in the form of box color selection, such as red for prediction boxes with a probability higher than 60%, and green for prediction boxes with a probability lower than 60%.

[0041] Through the above steps, the method for detecting Barrett's esophageal mucosal lesions disclosed in this invention can quickly analyze an esophageal image to be tested through the Barrett's mucosal detection model, and mark the lesion sites of Barrett's esophageal mucosa in the image in real time, as shown in Figure 4. It can also simultaneously visualize the risk or level of the lesion site, thereby achieving the effect of assisting judgment and improving the quality of medical care, and can also reduce the need for unnecessary biopsy examinations.

[0042] Furthermore, the method for detecting Barrett's esophageal mucosal lesions disclosed in this invention can be used in conjunction with an endoscope. This means that while the subject is undergoing esophageal endoscopy, the esophageal images can be analyzed in real time. This allows doctors to immediately see from the examination images whether the esophageal tissue has Barrett's esophageal mucosa, and thus make appropriate medical judgments and subsequent medical measures, thereby reducing medical time and preventing diseases.

[0043] The method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model disclosed in the second embodiment of the present invention is generally the same as that in the first embodiment, except that it further includes a step D. After step C, the detection result is input into the Barrett's esophageal mucosal dataset or the non-Barrett's esophageal mucosal dataset to expand the Barrett's esophageal mucosal dataset or the non-Barrett's esophageal mucosal dataset, thereby optimizing the performance of the Barrett's mucosal detection model. [Simplified Explanation of the Diagram]

[0044] Figure 1 is a flowchart illustrating the construction of a Barrett's esophageal mucosal detection model using an integrated object detection and classification model, as disclosed in the first embodiment of the present invention. Figure 2 is a flowchart illustrating the training, testing, and external verification of the object detection and classification model as disclosed in step A-2 of the first embodiment of the present invention. Figure 3 is a flowchart illustrating the object detection integration method in step A of the first embodiment of the present invention. Figure 4 shows the results of esophageal image analysis and prediction using the method of detecting Barrett's esophageal mucosal lesions assisted by the artificial intelligence model disclosed in the first embodiment of the present invention.

Claims

1. A method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model, comprising the following steps: Step A: Obtaining a Barrett's mucosal detection model; Step B: Obtaining an image of the esophagus to be tested; Step C: Inputting the image of the esophagus to be tested into the Barrett's mucosal detection model, wherein the Barrett's mucosal detection model analyzes the image of the esophagus to detect and locate areas in the image at risk of Barrett's esophageal mucosal lesions, thereby generating a detection result; if the Barrett's mucosal detection model detects areas in the image of the esophagus at risk of Barrett's esophageal mucosal lesions, the detection result includes a prediction box, used to identify areas in the image of the esophagus to be tested that are detected and determined to have Barrett's esophageal mucosal lesions, and to display the probability that the area selected by the prediction box is Barrett's esophageal mucosal lesions; wherein, The aforementioned actions of acquisition, input, analysis, and generation are each executed by a processing unit of a computing device according to program instructions.

2. The method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model as described in claim 1, wherein, The Barrett's mucosa detection model in step A is obtained through the following steps: Step A-1: ​​Obtain a Barrett's esophageal mucosa dataset and a non-Barrett's esophageal mucosa dataset, wherein the Barrett's esophageal mucosa dataset contains esophageal image data that has been determined to have Barrett's esophageal mucosa; and the non-Barrett's esophageal mucosa dataset contains esophageal image data that does not have Barrett's esophageal mucosa; Step A-2: Take at least one set of Barrett's esophageal mucosa dataset and at least one set of non-Barrett's esophageal mucosa dataset as a training dataset, input the training dataset into an integrated object detection and classification model for calculation and analysis, and produce the Barrett's mucosa detection model.

3. The method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model as described in claim 2, wherein, The esophageal imaging data in this non-Barrett's esophageal mucosa data set are all from the normal esophagus of a patient with non-Barrett's esophageal mucosa.

4. A method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model as described in claim 2, wherein, The esophageal imaging data includes an endoscopic image, a dynamic video, a histological image, and / or a static image.

5. A method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model as described in claim 2, wherein, The esophageal imaging data with Barrett's esophageal mucosa contained a lesion marker.

6. A method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model as described in claim 2, wherein, Step A-2 further includes using a separate dataset of Barrett's esophageal mucosa and a separate dataset of non-Barrett's esophageal mucosa as a test dataset, and inputting the test dataset into the Barrett's mucosa detection model for predictive analysis to confirm the predictive accuracy of the Barrett's mucosa detection model.

7. A method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model as described in claim 1, wherein, The esophageal image being tested is an endoscopic image located at the junction of the stomach and esophagus.

8. The method for detecting Barrett's esophageal mucosal lesions using an artificial intelligence model as described in claim 2, further comprising a step D, disposed after step C, wherein: Step D involves incorporating the prediction result into either the Barrett's esophageal mucosa dataset or the non-Barrett's esophageal mucosa dataset.

Images