Part I: Summary
Part II: Results
Part III: Conclusion and Future Work
GlioAI is an automatic brain cancer detection system that detects tumors in Head MRI scans.
Primary malignant brain tumors are the most deadly forms of cancer, partially due to the dismal prognosis, but also because of the direct consequences on decreased cognitive function and poor quality of life.
The issue with building an effective machine learning system for a medical imaging-based task before was that there was not enough data. Using transfer learning and implementing slight architecture modifications to adapt the model to our dataset allowed GlioAI to perform at over 95% accuracy.
Given the context of a transfer learning approach to a feature detection problem, it is crucial to safeguard the probability of model overfitting with prevention methods involving data augmentation, applying data normalization and dropout layer.
Magnetic Resonance Imaging is a new method that has emerged for improving safety in acquiring patient image data. The utilization of these imaging tools are not yet fully maximized due to the variable of human operation within detecting cancers without enough time to make an accurate prognosis.
Because manual image feature extraction methods are very time inefficient, limited to operator experience, and are prone to human error, a reliable and fully automatic classification method using MRI data is necessary for efficient cancer detection.
Reduce Mortality Rates
Controlling Treatment Output
Usability + Accessibility
We will be using a deep convolutional neural network, which is a neural network with a set of layers that will perform convolutions, pooling the set of regions of the image to extract features, along with with a softmax function that translates the last layer into a probability distribution.
We are only interested in applying transfer learning, which relies on training based on previously learned knowledge with a corpus of different large-scale datasets.
Because we are given a low volume of training data and are working with images, we decided to use VGG16, a state-of-the-art convolutional neural network with 16 layers to increase the probability of attaining a greater model accuracy.
The Brain MRI Images for Brain Tumor Detection was used to train the model which had 253 brain MRI scans.
The model was trained on 239 images belonging to two classes, and tested on 14.
The model consists of:
Model is trained on 25 epochs.
|No Transfer Learning||76%||49%|
When comparing the results of the different models that were trained, it is clear that the transfer-learning based model is the most accurate deep learning model to deploy for the web app.
Given that we can precisely automate the process of detecting whether a brain tumor is present in a patient or not, while simultaneously accompanying it with an easy-to-use user interface (for the doctor + patient), hospitals and patients will be able to simplify their workflow for detecting anomalies much earlier and are able to capture it with precision without having to sacrifice accuracy.
To further add, healthcare providers will be able to adjacently use applications that are built on top of the rapidly evolving tech infrastructure for care delivery with less friction of accessibility and utilization (via web).
There are many improvements to make within the models themselves to account for more diverse and unpredictable anomalies, which can be effectively improved in a cost-effective manner via generating more patient data to train the model using GANs.
After further model retuning and additional training optimization, GlioAI can specifically meet the pain points located within diagnosing brain tumors from MRI head scans, for brain cancer specialists and brain oncologists alike. Heading to a future where knowledge is aggregated and integrated with automated cancer detection systems in order to cut down diagnosis time over 1000-fold, from around 14 days of full reports to nearly 10-15 minutes, given the infrastructure for the crowdsourcing platform is built and incentive structures (via gig-based crypto token) and are aligned with verified physician users
In this coming decade (2020-2029), the necessity for automation within care delivery will hopefully be deployed at scale, putting the core central focus of the patient back into the hands of the care providers, while lining up monetary incentives for all parties involved via an inverse system between efficiency and cost with automation.
Given the current state that the model itself has been trained on a limited set(s) of patient MRI images with great accuracy, there is alot of area for improvement in terms of deploying extensive data augmentation (diversity of input image data for training), feature design, and overall application engineering and usability.
The future of GlioAI will be a web platform that will allow doctors to recieve feedback from other verified doctors in order to make a far more efficient and accurate diagnosis in less than half the time.
GlioAI, the collaborative encyclopedic medical platform for doctors, built for the 21st century.
The future of GlioAI lies in the idea of turning into a decentralized and pseudononymous crowdsourcing platform for medical practicioners and verified physicians and healthcare providers within the context of deep-knowledge tasks to further prune outputs from machines & automated systems ranging in disease detection and other areas in health.
Build a platform that can distribute high-value work (aggregation of feedback from board-certified doctors)
Game design mechanisms can be built out within the crowdsourcing platform in order to line up incentives for users to offer verified feedback that gets simultaneously ranked.
Propagandistic behaviors cannot occur because of the account verification process in order to create content or rank/upvote other posts (containment + authentication-based friction).
Enhancing treatment results via crowdsourcing platform specifically for verified doctors and healthcare providers (verified via school email + State ID)
Integrate gig-based cryptoeconomic mechanisms in order to incentivize (digital) teledoctors to be able to easily generate income via telemedicine tasks to ensure accuracy of diagnosis within timely conditions via providing direct emotional support + answer questions and make clarifications.
Crowdsourcing platform + machines = data-driven digital healthcare ecosystem
Contributions are always welcome! For bug reports or requests please submit an issue.