MeAI tool for healthcare system

Compnay Name

Service Details

Medical AI (MeAI) provides a primary API function for users to obtain diagnostic results. Users can upload medical images, such as endoscopic images, as input to a trained AI model. MeAI will perform inference in order to provide diagnostic results by utilizing the feature extraction process of complex neural network architectures. These results provide information regarding the likelihood and kind of disease the user may experience. Due to the use of sophisticated AI techniques and thorough training, obtaining a user's diagnosis not only results in high accuracy, but also a significant reduction in waiting time.

Problem Description

According to WHO, cancer is leading cause of death in the world and we predict the number of cases to increase over the next decade. Early detection of cancers and effective treatment can reduce death rates. As for stomach cancers, early diagnosis of polyps and lesions gives patients the good chance to have successful care and treatment and increase their lifetime. Today, many people suffer from stomach diseases because of work pressure and unhealthy lifestyle such as staying up late, drinking alcohol. According to

in 2020, stomach cancer ranks the fifth most common cancers worldwide and is fouthe fourthading cause of cancer deaths. In Vietnam, according to data from the Ministry of Health, there are about 26% of the population has stomach diseases. Seriously, the age of gastric cancer patients is getting younger and younger and the rate of patients under 40 years old being

.

One of the feasible methods for early diagnosis of many types of cancers is using Endoscopic. Traditional methods of endoscopic rely on manual inspection by healthcare experts, and therefore is affected by various factors such as the work environment, lighting, equipment quality, and the expertise of the physician. This leads to a relatively high rates of misdiagnosis and large number of patients cannot have access to early detection and timely treatment. When the number of patients is too large, doctors are overworked and become tired, leading to a higher rate of omissions and misdiagnosis. Recent studies record the rate of misdiagnosis for patients undergoing endoscopy is

and this rate seems to be higher in the medical centres that are overloaded or poorly equipped. The endoscopy misdiagnosis is even more severe for low-income people and those in non-central areas who are unable to undergo routine cancer screenings or repeat check-ups at the reputable hospitals.

Aim to assist people to recheck their diagnosis results, our project builds an online tool that enable people to submit their stomach endoscopic photos and advise them on the probability of having polyps, gastritis areas and cancers. Based on the given results, people have more information to decide whether to go for a re-examination at another healthcare system.

Solution Description

Our solution is to build an AI model for detecting the abnormal area in the stomach endoscopy images. There are four steps of this process.

1. Analyzing problems

We work with several Specializations in Gastroenterology to describe the characters of the stomach endoscopy image which have polyps, gastritis areas and cancers. From the description of the experts, we will use some data analytic tools to extract the feature of interested (ROI) areas of image data for deeper understanding the type of injury of stomach. With this knowledge, we will draw AI solutions to detect the injury areas of the stomach endoscopy images more effectively.

2. Collecting and labeling data

The data is the "food" of the AI models. It has very high impact to the performance of the AI models for this problem. Thus, collecting and labeling data are very important tasks. We may need to outsource this tasks for the experts in the Gastroenterology field. In the first version of this tool, we will collect the public data for building the AI models. The data is verified by the experts to enhance the clean data. This will be done in the M1 milestone of our project.

3. R&D AI models

With the deep understanding about the problem and the data, we do the R&D task for researching the state-of-the-art object detection models. They consist of, but not limited to, the following models:

- CenterNet

- CornetNet

- Faster R-CNN

- Mask R-CNN

- Yolo

- RentinaNet

4. Training AI models

With the selected AI models, we do training with our clean data. Here, we will use some strategies for training good AI models. We also need to hire the resource for training AI models with a large number of image data. After the training process, we will have several AI models. Thus, we need to choose the best model based on the testing data for deploying. These all tasks need to be executed several times until achieve the best model.

5. Deploying AI model to the service When we have AI model, we need to deploy to the MeAI service. I drew the description of our service in the next section. We plan to finish this tool in the M3 milestone of the project timeline.

6. Collecting data for further improvement

To further improvement, we will collect the user data with their acceptance when using our tool. The data will be used to update the AI models for enhancing the performance in the future.

Milestone & Budget

Revenue Sharing

Related Links

[1]

Long Description

Company Name

Cardano2vn.io Company

I. Summary

The Medical AI (MeAI) tool is an online tool that enables people who have had a gastroscopy to recheck the diagnosis results based on their endoscopic photos. All users need to do is input their endoscopic images. The tool integrated with a smart medical AI system will give users advise of whether they have polyps, gastritis areas, or cancerous lesions. Users then decide if they should go for another endoscopy. This tool is great valuable for low-income people and those living far from cities who cannot afford to go for regular check-ups or re-examination at the reputable medical centres.

II. Funding Amount

$ 39,000

III. The Problem to be Solved

According to WHO, cancer is leading cause of death in the world and we predict the number of cases to increase over the next decade. Early detection of cancers and effective treatment can reduce death rates. As for stomach cancers, early diagnosis of polyps and lesions gives patients the good chance to have successful care and treatment and increase their lifetime. Today, many people suffer from stomach diseases because of work pressure and unhealthy lifestyle such as staying up late, drinking alcohol. According to

in 2020, stomach cancer ranks the fifth most common cancers worldwide and is fouthe fourthading cause of cancer deaths. In Vietnam, according to data from the Ministry of Health, there are about 26% of the population has stomach diseases. Seriously, the age of gastric cancer patients is getting younger and younger and the rate of patients under 40 years old being

.

One of the feasible methods for early diagnosis of many types of cancers is using Endoscopic. Traditional methods of endoscopic rely on manual inspection by healthcare experts, and therefore is affected by various factors such as the work environment, lighting, equipment quality, and the expertise of the physician. This leads to a relatively high rates of misdiagnosis and large number of patients cannot have access to early detection and timely treatment. When the number of patients is too large, doctors are overworked and become tired, leading to a higher rate of omissions and misdiagnosis. Recent studies record the rate of misdiagnosis for patients undergoing endoscopy is

and this rate seems to be higher in the medical centres that are overloaded or poorly equipped. The endoscopy misdiagnosis is even more severe for low-income people and those in non-central areas who are unable to undergo routine cancer screenings or repeat check-ups at the reputable hospitals.

Aim to assist people to recheck their diagnosis results, our project builds an online tool that enable people to submit their stomach endoscopic photos and advise them on the probability of having polyps, gastritis areas and cancers. Based on the given results, people have more information to decide whether to go for a re-examination at another healthcare system.

IV. Our Solution

Our solution is to build an AI model for detecting the abnormal area in the stomach endoscopy images. There are four steps of this process.

1. Analyzing problems

We work with several Specializations in Gastroenterology to describe the characters of the stomach endoscopy image which have polyps, gastritis areas and cancers. From the description of the experts, we will use some data analytic tools to extract the feature of interested (ROI) areas of image data for deeper understanding the type of injury of stomach. With this knowledge, we will draw AI solutions to detect the injury areas of the stomach endoscopy images more effectively.

2. Collecting and labeling data

The data is the "food" of the AI models. It has very high impact to the performance of the AI models for this problem. Thus, collecting and labeling data are very important tasks. We may need to outsource this tasks for the experts in the Gastroenterology field. In the first version of this tool, we will collect the public data for building the AI models. The data is verified by the experts to enhance the clean data. This will be done in the M1 milestone of our project.

3. R&D AI models

With the deep understanding about the problem and the data, we do the R&D task for researching the state-of-the-art object detection models. They consist of, but not limited to, the following models:

- CenterNet

- CornetNet

- Faster R-CNN

- Mask R-CNN

- Yolo

- RentinaNet

4. Training AI models

With the selected AI models, we do training with our clean data. Here, we will use some strategies for training good AI models. We also need to hire the resource for training AI models with a large number of image data. After the training process, we will have several AI models. Thus, we need to choose the best model based on the testing data for deploying. These all tasks need to be executed several times until achieve the best model.

5. Deploying AI model to the service When we have AI model, we need to deploy to the MeAI service. I drew the description of our service in the next section. We plan to finish this tool in the M3 milestone of the project timeline.

6. Collecting data for further improvement

To further improvement, we will collect the user data with their acceptance when using our tool. The data will be used to update the AI models for enhancing the performance in the future.

V. Service Description

Medical AI (MeAI) provides a primary API function for users to obtain diagnostic results. Users can upload medical images, such as endoscopic images, as input to a trained AI model. MeAI will perform inference in order to provide diagnostic results by utilizing the feature extraction process of complex neural network architectures. These results provide information regarding the likelihood and kind of disease the user may experience. Due to the use of sophisticated AI techniques and thorough training, obtaining a user's diagnosis not only results in high accuracy, but also a significant reduction in waiting time.

VI. Our Project Milestones and Cost Breakdown

The proposed budget for the MeAI tool is $39,000. This

is the overall project milestones and cost breakdown.

VII. Risk and Mitigation

The construction and implementation of this MeAI system contains several risks that might affect the success of our project. We list below some potential risks and suggest solutions to minimize their impacts to the project.

1. Data Collection

Risk: This is a specific problem based on gastric endoscopy images. The collection of data for the project is relatively difficult as gastric endoscopy image datasets are rare and almost unavailable to public .

Mitigation solution: We will make a detail plan of conducting surveys and collecting data in different provinces and cities in Vietnam. 2. Data accuracy

Risk: Inaccurate labelling of input endoscopic images will affect the diagnosis model results. Endoscopic images are collected from patients and the medical examination results of these might be incorrect as a result of endoscopy misdiagnosis.

Mitigation solution: We will work closely with medical experts to check the accuracy and consistency of the data.

3. Effectiveness of the diagnosis model.

Risk: The diagnostic model is not effective

Mitigation solution: We strengthen research and deployment of AI models that solve similar problems. Experiments are carefully designed and conducted to evaluated models in different scenarios. These will give rigorous guide in selecting a right model.

4. Effectiveness of the diagnosis model.

Risk: The diagnostic model is not effective.

Mitigation solution: We strengthen research and deployment of AI models that solve similar problems. Experiments are carefully designed and conducted to evaluated models in different scenarios. These will give rigorous guide in selecting a right model.

5. Model resources

Risk: Most AI models face difficulties in model resources.

Mitigation solution: We will hire cloud service for training and evaluating models

6. Response time

Risk: Most AI models need large resources. However, the lack of GPU and resources to deploy the model on the SNET platform leads to the slow response of the models.

Mitigation solution: We will research and deploy technologies to increase the performance of AI models such as using deepstream, GRPC and tensorRT.

7. User experience

Risk: It is required that users need to use computers and smartphone relatively well and correctly follow instructions.

Mitigation solution: We build a user-friendly, easy-to-use and highly robust system and attach detailed, easy-to-understand user guide and video instructions.

VIII. Voluntary Revenue

There are two main sources such as Charge based on Request and In-App Advertising can be the voluntary revenue generation of the MeAI tool.

IX. Our Team

Ly Vu - Product Lead, Modeling and AI R&D

• Senior Data Scientist / AI Researcher
• Lead AI Core Developer and Maintainer
• Former Research Scientist with Anomaly detection
• 9+ years in the computer vision, natural language processing domains
• 3+ years of Python experience to build and develop AI models (including eKYC products, information extraction products
  )
• Google scholar
  link

Van Pham - Data analytics, Testing

• Business analytic
• Data analytic
• Data collection and labeling
• Testing and Analysing errors of AI models
• Link

Phong Tran - Software Engineering and Integration

• Senior Software Engineer at VNG cooperation, Vietnam
• 3+ years of experience in AWS cloud development
• 2+ years of experience in web application development

Anh Tien Nguyen - Software Engineering and Software management

• Senior Software Engineer at several banks in Vietnam

X. References

[1]

AI Services