Image classifier differentiating tumour from healthy MRI scans using convolutional neural networks

Jessica S. Velasco; August C. Thio-ac; Gilbert B. Hollman; Benedicto N. Fortalezaa; Felix Kenjie F. Thio- ac; Jonard Bryan S. Cabanding; Jetlee E. Dumandan; Lean Karlo S. Tolentino; Roel M. Mendoza

Jessica S. Velasco Department of Electronics Engineering, Technological University of the Philippines,Manila,Philippines

August C. Thio-ac Department of Electronics Engineering, Technological University of the Philippines,Manila,Philippines

Gilbert B. Hollman Department of Electrical Engineering, Technological University of the Philippines, Manila, Philippines

Benedicto N. Fortalezaa Center for Engineering Design, Fabrication, and Innovation, College of Engineering, Technological University of the Philippines, Manila, Philippines; Department of Electronics Engineering, Technological University of the Philippines, Manila, Philippines

Felix Kenjie F. Thio- ac Department of Electronics Engineering, Technological University of the Philippines,Manila,Philippines

Jonard Bryan S. Cabanding Department of Electronics Engineering, Technological University of the Philippines,Manila,Philippines

Jetlee E. Dumandan University Extension Services Office, Technological University of the Philippines, Manila, Philippines; Department of Electronics Engineering, Technological University of the Philippines, Manila, Philippines

Lean Karlo S. Tolentino Department of Electronics Engineering, Technological University of the Philippines,Manila,Philippines

Roel M. Mendoza Department of Electrical Engineering, Technological University of the Philippines, Manila, Philippines

This study was created using a simple algorithm that may be implemented on MRI scan machines to aid medical professionals in making a more accurate decision by having a second opinion. The system is inspired on a smartphonebased skin disease classifier which was implemented with multiple techniques of data augmentation. It is also trained on multiple pre-trained and untrained algorithms for machine learning application. The raw data of MRI has a dimension of 50 x 50 and in colored form. When worked on raw data it did not yield great results having only an 80% accuracy and a 50% validation accuracy after. We implemented data augmentation techniques to increase the accuracy and to implement a reasonable comparison among the different CNN models. The data was resized and recolored to a monochrome to have the same dimensions and tone to have a better prediction on the data. These images were then trained on a well-known pretrained Convolutional Neural Network (CNN) and a small model comprising of 3 layers for control independent variable. At first, we only got 90% as the best validation accuracy that belongs to the untrained 3-layer algorithm. However, decreasing learning rate to 0.0001 increased our best validation accuracy from 90% to 95% that validation accuracy belonging to VGG 16 and 19.