Early detection and segmentation of malignant brain tumors using machine learning

This project aims to develop fast, efficient and reliable methods for diagnosing brain cancer by deploying computers for the task.   Cancers of the brain require an ability to discern patterns in large, noisy datasets in order to be diagnosed. Computers can utilize techniques to analyze

Project Title

Early detection and segmentation of malignant brain tumors using machine learning

Project Area of Specialization

Artificial Intelligence

Project Summary

This project aims to develop fast, efficient and reliable methods for diagnosing brain cancer by deploying computers for the task.  

Cancers of the brain require an ability to discern patterns in large, noisy datasets in order to be diagnosed. Computers can utilize techniques to analyze extremely large volumes of data almost instantaneously. Moreover, computers can use those techniques to deliver accurate results when their human counterparts are unable to do so. These techniques fall under the discipline of deep learning.

Deep learning uses sophisticated math to generate models that can learn how to spot particular patterns in unseen data. Existing literature shows that using deep learning for diagnosis of cancer can result in earlier detection of malignant tumors and consequently better survival rates for patients. And while it doesn't eliminate the need for medical professionals, it can help rectify the global paucity of neuro-oncologists

Project Objectives

Short-term objectives:

1. Develop a machine learning model that can detect and isolate tumorous growth in the human brain with high accuracy and reduced latency, using Magnetic Resonance Imaging (MRI) scans supplied by the user.

2. Predict whether an identified tumor is benign or cancerous by parsing the MRI scans down to individual layers and analyzing those layers for biomarkers particular to brain cancer.  

3. Use the provided MRI scans coupled with non-identifiable information collected from the user to stratify risk profiles for patients.

4. Provide precise segmentation of a tumorous region to help doctors plan cancer treatments, such as surgery or radiation therapy, without damaging neighbouring healthy tissues.

5. Train our model to also spot early indicators of tumorous growth, such as inflammation and lesions, on MRI scans.

6. Develop a user-friendly, machine specification independent web interface for interacting with our model that is accessible to anyone, anywhere.

Long-term objectives:

1. Develop and evaluate novel biomarkers for early cancer detection and patient risk profile stratification by analyzing the data accumulated through our app.

2. Collaborate with experts from multiple disciplines to help speed up the process of brain cancer diagnosis by translating novel research techniques and discoveries into viable diagnostic methods by inculcating them in our model.

3. Analyze collected data to better understand behaviours related to screening uptake in different segments of the population.

4. Extend our model to facilitate detection of brain abnormalities other than cancer such as cysts and deformities.

Project Implementation Method

In this project we are focusing on building a tumor detection model using a convolutional neural network in TensorFlow & Keras.

There are two phases of operation:

• Pre-processing  
• Segmentation.  

Skull stripping and image enhancement methods are used in pre-processing. The MRI images are converted into grey scale images which are then smoothened by adjusting the contrast.

Convolutional Neural Network (CNN) is used for segmentation of tumors.

Loss function is optimized using the BAT algorithm.

Benefits of the Project

• According to the World Cancer Research Fund, 0.5 million cases of brain cancer are reported every year. 64% of these cases go on to prove fatal in about 5 years. Early diagnosis of brain cancer can help bring down this percentage to about 30%.

• Detecting brain cancer on MRIs is notoriously difficult. As a result, 30% of cancerous brain tumors have metastasized, i.e. spread to other parts of the brain, before diagnosis. Deferring the task of diagnosis to computers makes it not only faster and more efficient, but also less error prone and more reliable.

• Earlier diagnosis of cancer offers the greatest potential for transformational improvements in patient outcomes. A patient diagnosed with stage 1 brain cancer has over 70% chance of survival beyond one year. This drops to less than 15% if diagnosed at stage 4.

• For up to 60% of patients, headaches are the only symptoms that they experience. Going to an oncologist for a headache seems like an unnecessary hassle which is why most cases of brain cancer are diagnosed at least a year too late. Computers can simplify the tedious process of initial diagnosis and discovery and, as a result, make earlier treatment of brain cancer a possibility.

• Brain cancer has one of the lowest rates of survival of all cancers in Pakistan. This is because of the absolute dearth of facilities for brain cancer in the country. Particularly, there are only 2 centers of neuro-oncology in Pakistan and only 1 oncologist for children. This makes the already abysmal rates of survival for this cancer even worse. A computer aided solution can help rectify this shortage.

Technical Details of Final Deliverable

The final deliverables for our project include:

1. A trained convolutional neural network (CNN) with an appropriate data cleaning and preparation pipeline. This pipeline will be able to process and correct images fed to it and amplify contained markers that will aid the detection process. The CNN will then parse the output of this pipeline to discern between the different strata of the cerebral tissue, segment extraneous growths within those strata and then classify those growths as malignant or benign.
2. A user-friendly interface that lets people without the ability to operate highly technical systems interact with our model. This interface will be hosted on the web so that it is largely independent of the user's computer specification. This is because neural networks require high end hardware in order to function. This is also to facilitate ubiquitous access to the model. The interface would be connected to the model on the backend. On the frontend, it would take in images of MRI scans from the user and direct those images to the backend. From there, the data will make its way through the trained model, which will then emit an analysis for the supplied MRI.

Final Deliverable of the Project

Software System

Core Industry

Medical

Other Industries

Core Technology

Artificial Intelligence(AI)

Other Technologies

Sustainable Development Goals

Good Health and Well-Being for People

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com