Novel Convolutional Neural Networks for Improved Accuracy in User-accessible Brain Tumor Detection
Session Number
RISE 09
Advisor(s)
Mrs. Allison Hennings, Illinois Mathematics and Science Academy
Mr. Sean Fu, Tesla
Mr. Thomas Walton, Georgia Tech
Dr. Haohan Wang (Ph.D.), Carnegie Mellon University
Discipline
Medical and Health Sciences
Start Date
17-4-2024 10:45 AM
End Date
17-4-2024 11:00 AM
Abstract
The purpose of this design investigation was to create three user-friendly artificial intelligence (AI) models, specifically convolutional neural networks (CNNs), each designed using different and novel techniques for brain tumor detection from Magnetic Resonance Imaging (MRI) scans. These models aimed to assist medical professionals in overcoming inefficiencies, reducing human error, and surpassing existing models in accuracy by addressing various gaps and limitations in their work. The design began by developing three CNNs, each employing different techniques which were then trained on a preprocessed dataset containing MRI scans of healthy and tumorous brains. The third model, utilizing the pretrained VGG19 architecture, performed the best and was then integrated into a user-friendly website allowing hospital workers to upload MRI images for tumor identification. A CNN that could discern between different tumor types and malignancy levels was desired but unable to be developed. In conclusion, a user interface was developed featuring the superior CNN model. This model, which had the highest accuracy (99%) and lowest loss (0.01), and an F-score of 98, was able to address some of the gaps in previous models, including low accuracy, the use of generic architectures, and the lack of a user interface. It took the model mere seconds to predict the outcome of a scan, providing a significantly faster alternative to manual tumor detection which could take hours or even days.
Novel Convolutional Neural Networks for Improved Accuracy in User-accessible Brain Tumor Detection
The purpose of this design investigation was to create three user-friendly artificial intelligence (AI) models, specifically convolutional neural networks (CNNs), each designed using different and novel techniques for brain tumor detection from Magnetic Resonance Imaging (MRI) scans. These models aimed to assist medical professionals in overcoming inefficiencies, reducing human error, and surpassing existing models in accuracy by addressing various gaps and limitations in their work. The design began by developing three CNNs, each employing different techniques which were then trained on a preprocessed dataset containing MRI scans of healthy and tumorous brains. The third model, utilizing the pretrained VGG19 architecture, performed the best and was then integrated into a user-friendly website allowing hospital workers to upload MRI images for tumor identification. A CNN that could discern between different tumor types and malignancy levels was desired but unable to be developed. In conclusion, a user interface was developed featuring the superior CNN model. This model, which had the highest accuracy (99%) and lowest loss (0.01), and an F-score of 98, was able to address some of the gaps in previous models, including low accuracy, the use of generic architectures, and the lack of a user interface. It took the model mere seconds to predict the outcome of a scan, providing a significantly faster alternative to manual tumor detection which could take hours or even days.