Abstract:
The lack of radiological equipment/personnel and difficulties in acquiring licensure for
MRIs and CT scans in provinces in far flung areas of the Philippines remain to be a
hindrance in timely medical intervention. Additionally, development of a non-invasive
tool for initial diagnosis and classification of brain tumors is of interest due to its high
prevalence. CNN-based MRI image analysis aims to classify between meningioma and
glioblastoma brain tumor types and between the presence and absence of brain tumors.
The study obtained glioblastoma, meningioma, and normal brain MRIs from publicly
available online databases for training, verification, and validation. After training,
comparisons of model loss and accuracy were done with Adam, SGD, RMSProp, and
Adadelta. SGD was found to be the most effective optimizer in normal and brain tumor
classification (validation accuracy of 87.00%) while RMSProp was the most effective for
normal, meningioma, and glioblastoma classification (validation accuracy of 96.33%).
Various factors such as increased image dataset size, number of subjects, number of slices
for each category, and types of brain tumor is recommended for future studies.
Furthermore, the investigators recommend including images from Philippine hospitals in
training classification while setting ample training time as pre-trained models would
benefit from a longer training time.
