AI Product Business Proposal: Early Detection of Breast Cancer
The sample business proposal for early detection of Breast Cancer which includes defining business goal, success metric, data collection, model selection and eventually deploying model.
Business Goal
A business problem
In Pakistan, it’s getting harder and harder to detect breast cancer due to social factors and the availability of radiologists, even then this process requires focus and time which is a luxury if you are in the middle of a pandemic. The radiologist takes a mammogram and then detect tumor or cancer with the naked eye which can be difficult sometimes and error rates are quite high if you are under stress.
Out of one thousand women, about one hundred are recalled for additional diagnostic imaging, and of these one hundred women, four or five are diagnosed with breast cancer (nih.gov) (See Figure 1). These false-positive exams lead to preventable harms, including patient anxiety, benign biopsies, and unnecessary intervention or treatment. Furthermore, high false-positive rates significantly contribute to the annual $7.8 billion mammography screening costs (oup.com). These problems can be solved by emerging Artificial Intelligence (AI) based computer vision technologies, which will reduce the rate of false-positive prediction and improve the efficiency of the radiology department.
Get Abid Ali Awan’s stories in your inbox
Join Medium for free to get updates from this writer.
The business solution is to improve the effectiveness of radiologists by making the detection part smooth so that they can focus on providing accurate and faster results.
Business case for the product
This product will increase patient satisfaction as they don’t have to wait 2 days to get the report and even then, the report will be accurate enough that they don’t have to go for another mammogram. This will increase customers’ trust and increase revenue for medical labs and hospitals, directly and indirectly through customers satisfaction. The machine learning will reduce the time to process the images and predict whether it is benign or malignant. Which is quite difficult to detect from the naked eye? This product will help doctors to determine whether the patient needs immediate surgery or need chemotherapy.
AI/ML model
This can be done by using Deep Learning Computer Vision, in which we will train our model on X-ray images of breasts, with cancer and without cancer. The introduction of the Machine learning model will improve the performance of the radiology department in each hospital, and this will directly improve the lives of the radiologist who are working continuously.
Success Metrics
The success metric will be measured by monitoring the increase in revenue of the radiology department and customer satisfaction. The customers include radiologists, patients, and oncologists. The success metric can also be measured by the rating given by a radiologist or oncologist. We want to improve the accuracy of cancer detection and reduce the time to observer a mammogram, which indirectly will help oncologist makes decision faster, even in the pandemic outbreak.
Data
Size and potential biases in the data
Initially, we are going to get five hundred images of mammograms and label them either in-house or hire a third party. We will use Appen to monitor and improve image labeling. At first, we will be going for only two categories. With Cancer or without cancers and we will make them balance to avoid biased prediction.
How the data will be acquired?
We will collect some of the images from hospitals and publicly available data sets and rest of the data we will either buy or sign MOUs with different hospitals so that we can easily get access to new datasets from the real world and in return, we will provide the service at a discounted rate.
Data labels
Initially, we will keep it simple with only two categories and with time we will add up to six categories. We are also planning to add three subcategories of benign, malignant, and unknown. It will be easy to deploy and maintain two categories of the dataset and the model performance will also increase on binary tasks. The simple solution will require a radiologist or oncologist to review the photo again to determine the type of cancer and then determine whether it’s benign or malignant.
Model
How model will be built?
Due to privacy concerns and sensitivity of data we want to keep the model private and train it in-house, but due to lack of resources, we will be forced to use cloud service at the initial stage. We will train our model on the cloud and create inference on the cloud. This will help us launch a minimal viable product, but in the second stage, we want to move both model and dataset from cloud to local. We will train the dataset on various higher tiers models such as CNN and RNN and then ensemble them to get the best results. This process is iterative and with more that’s, we will keep updating our model with time.
ML metrics
In cancer detection, recall and precision is the most important metric that we want to improve and subsequently improve overall effectiveness by monitoring the f1 score. The accuracy will be used for marketing purposes but for model performance we will stick to recall and f1 score. To learn more on metrics click here.
Minimum Viable Product (MVP)
User interfaces
The initial interface is built on the cloud and has a simple interface. First, it will ask you to add an image of a mammogram and then it will process the image and after a few seconds, it will give you the result of whether this image contains cancer or not. It will also give confidence score and process time.
Types of users
The interface will best user-friendly and it’s almost similar to the UI that they are used to in medical professions. We will keep it simple at first so that they don’t need to take training or look up to user manual.
Building and launching the product
We will start small by training the model on two categories and improve our f1 score by preprocessing and model selection. At first, this will be done on Google cloud but with time the entire process will be shifted to the local base. At the end of every result, the UI will ask the user to put a rating on our results and how fast we got the result. This will help us determine success metrics. Overall, the whole product will be launched on cloud services, and users either will be able to connect the database or upload images individually.
Post-MVP-Deployment
Improve the product in the long term
- We will monitor the model performance metric and try to improve by training it on a new batch of images.
- We will monitor success metrics and customer reviews on our products to provide them with a better interface and product performance.
- We will be adding more categories of cancer and further add whether it is benign or malignant.
- We will keep on improving the interface so that it will take a few clicks to get the results that the doctor wants.
- We will also improve the quality of image processing with time to give us better results.
- Active learning will be the main part of this product as we will be getting new data from patients.
- We will improve the model performance for unknown category by adding humans in the loop such as radiologist or oncologist.
Monitoring bias
To mitigate bias, we will be collecting data from various hospitals and try to keep our database balanced. This will be done under active learning so our product can predict accurately a new type of tumor. We will also keep a human in the loop to monitor the performance of our product, most probably in house oncologist, who will monitor the unknown category and try to label the data to improve the performance.
Motivation
In Pakistan, we have multiple startups in artificial intelligence especially in medical sector such as XYLEXA who are building products using computer visions such as CADx which detect breast cancer with 93% accuracy. These startups need guidelines to get into mainstream market and start selling product.
Secondly, Pakistan is facing shortage in biomedical staff and certified radiologist, which get worst during pandemic. When qualified staff are found, they are often lost to ‘brain drain’. Brain drain refers to educated workers emigrating from their developing world nations (nih.gov).
Conclusion
AI is quickly allowing developing countries to solve pressing challenges in critical sectors like health, energy, agriculture, and financial services (worldbank.org). Especially in Pakistan, the health care industry needs AI technology to improve the effectiveness in diagnosing the disease and provide faster results so that specialists can act quickly to save patient’s life. We have introduced the sample business plan for AI product which provide a clear view on how we can introduce a new product in any company. The business goals include, business benefit of proposed solution, the data gathering and AI models selection. Then, we have discussed success metrics and data collection methods. At last, we focused on model selection, training and then deploying it as a minimum viable product with a simple user interface. We have also provided clear direction on how we will be maintaining and improving our AI application for post deployment versions.
I hope this article has helped you convert AI application Ideas into reality.
