Detecting Brain Tumors in Medical Images with Deep Learning

Authors

  • Ahmed Samit Hatem College of Nursing, University of Kerbala, Kerbala, Iraq https://orcid.org/0000-0003-0465-1866
  • Hussein Ali Mahdi College of Nursing, University of Kerbala, Kerbala, Iraq.

DOI:

https://doi.org/10.71229/gjsz0k21

Keywords:

Brain Tumors, Deep Learning, EfficientNet, InceptionV3 , TTA and ensembles

Abstract

Brain tumors are characterized as abnormal growths of cells inside the brain. They can either be benign or malignant; therefore, an earlier and more accurate diagnostic process will enhance better outcomes in patients affected by this disease. Magnetic Resonance Imaging (MRI) is a primary modality used to detect the existence of such anomalies, but since its interpretation requires manual analysis, it takes up a lot of time and has been found to lead to errors due to involvement of humans. The advent of deep learning (DL) techniques has made it possible to introduce greater accuracy in determining the nature of such tumorous masses, with increased consistency. A novel classification framework combining the advantages of both transfer learning and two-stage fine-tuning, test-time augmentation (TTA), and finally, ensemble learning is proposed in this study, enabling classification of brain MRIs into four categories – glioma, meningioma, pituitary tumor, and no tumor. Two families of pretrained ImageNet convolutional neural networks (CNNs): EfficientNet (versions B3 and V2S) and InceptionV3 are tested against the Brain Tumor dataset consisting of 2723 images. For this purpose, we have applied several data augmentation techniques as described earlier to ensure that data would be enriched and avoid overfitting on the training set. All the models are trained under the same two stage approach: the first train the classification head on top of a fixed backbone, then fine tune it only on the last 150 layers. Then we built ensembles of three models of each kind, varying the random seeds or architecture for the latter.  As a result, when tested against the 545-image dataset, the Accuracy score achieved by an ensemble based on EfficientNet was of 96.51% (with a Macro-F1 score of 96.4) while the InceptionV3-based one scored higher at 96.88% Accuracy and 97 Macro-F1 score respectively. Although the results are slightly more impressive for the InceptionV3 model, the differences between the two ensembles are minor. Both types work best for detecting pituitary tumors and find detecting meningioma ones rather challenging.

fig 9

Downloads

Published

2026-10-15

Issue

Section

Original Articles

How to Cite

Detecting Brain Tumors in Medical Images with Deep Learning. (2026). Al-Noor Journal of Engineering Management and Computer Science, 2(4), 586-596. https://doi.org/10.71229/gjsz0k21

Similar Articles

51-60 of 66

You may also start an advanced similarity search for this article.

Most read articles by the same author(s)