Best Projects 2021 features all the nominated entries submitted ‘Submit your Project’ category. All the entries consist of innovative projects run by existing enterprises in the form of businesses, NGOs or informal programs.
A Novel Mask R-CNN Model to Segment Heterogeneous Brain Tumors through Image Subtraction
Published on June 1, 2021 um 09:02
Type of Enterprise
Year the ngo or company was founded
Explain your project in details:
The goal of the project is to train a Mask R-CNN to segment brain tumors, mostly Gliomas, using image subtraction on MRI scans. Brain tumors affect thousands and there are few cures. Brain surgery is often required, but unfortunately has a 22% success rate. The perfect segmentation of the tumor is vital for a successful surgery. For this reason, I wanted to create a model that can locate tumors to help radiologists segment with better success. I researched methods that I could implement including both the medical and CS field, coming across a method used in the radiological field called image subtraction. Radiologists often take four types of MRI scans of the patient before analyzing the scans (2: enhanced, 2: non-enhanced). They subtract the non-enhanced image from the enhanced image and use the resulting output when trying to segment the tumor. I incorporated this medical method to train Mask R-CNN. I believe that this method will also be helpful to models since it provides more information. When performing background research with existing brain tumor segmentation models, I came across many types of architectures and techniques, I have yet to see one using image subtraction. Both networks had higher performance when using image subtraction instead of normal MRI scans. If this method or idea be applied to current state-of-the-art brain tumor segmentation models, this could really improve the computer science models and potentially be more beneficial to the medical aspect.
Impact of your enterprise on sustainable development
I was able to prove that image subtraction helped through this project, receiving a final f1 score of 0.63 vs 0.68 on the T1 MRI scan network and 0.65 vs. 0.71 on the T2 MRI scan network. In terms of general comparison terms, the model, even though trained for only a few epochs, got 0.03 less than the state of the art models. If this method were to be applied to that model, it would result in even better performance. Based on results, I conclude that richer information of image subtraction helps models learn better features and hence can predict with greater accuracy. It showed improvement in all measured metrics as shown in the table. Mean in the table means average across all pixels. This is a common metric used in segmentation problems. This algorithm can be implemented and used in various ways. Specifically, this model can help doctors and radiologists segment brain tumors with higher accuracy. It should be kept in mind that this algorithm is meant to be a tool at health professional's disposal rather than a replacement. This algorithm could be used in other methods of Mask RCNN in the medical field where image subtraction is possible.
Sustainability and future plans
While this project has specific requirements, this model can be expanded to the other depths of healthcare. For example, the main purpose of this project was to be able to segment brain tumors for better successful surgery. There are many other diseases such as blood clots that require brain surgery. There are also many forms of x-rays taken for diseases and infections that occur outside the brain. The same method(s) can be applied to these other diseases with some form of segmenting location that can expand this product/project. In terms of selling and financing source. This product will be sold to hospitals rather than the general public since it can be misused. Currently, I am talking to a consultant radiologist in london who agrees that this product could be quite beneficial since it allows for radiologists to use a second hand while not hiring any other radiologists and making the final decision.
Your profile as an entrepreneur
I am currently a high school student (14) at BASIS Independent Silicon Valley. I have a passion for machine learning which I express by participating in science fairs and presenting research at conferences. Growing up with a tech engineer as a father and an art teacher as a mom, I was introduced to technology with a viewpoint of improving items to make them more accessible and practical. In fact, in 2019, I received Project on the Year at California State Science Fair for a computer vision project and in 2020 received the top 300 projects in Broadcom award, as well as recently published a preprint on Arxiv. In 2021 I received honorable mention in NCWIT. I also received 1st in the Synopsys Science fair in 2021 for another machine learning project. Growing up, I noticed the lack of resources provided to students in deeper subjects, such as AI. While science, math, and English were all electives given to middle and high school students, fewer programming and CS electives were given as options. Artificial intelligence will be our future ("AI is the new electricity" - Andrew Ng) and everyone should have the option to at least be introduced to the field. For this reason, I created AndromedaAI, a non-profit organization, meant to be a platform for students like myself to learn about artificial intelligence as well as compete in monthly machine learning challenges for top prizes. I also recently joined the StartOnAI non profit organization, which aims to help spread the learning of AI through books, tutorials, videos, etc. Outside of my general passion for artificial intelligence, I love to hang out with friends and play club volleyball at City Beach. I also like to play the ukulele in my spare time as well as listen to music.
Segmentation of brain tumors through MRI scans using MRCNN using a method of image subtraction
Brain Tumor Project Board