DenseNet201-Based Deep Transfer Learning Framework for Brain
Tumor Classification in MRI Scans
Doaa Sami Khafaga1,*
1Department of Computer Sciences, College of Computer and Information Sciences, Princess Nourah bint
Abdulrahman University, P.O. Box 84428, Riyadh 11671, Saudi Arabia
Email: dskhafga@pnu.edu.sa
Abstract
The classification of brain tumors is crucial in the context of early intervention, as the appropriate and
timely diagnosis can significantly influence the treatment plan and patient outcomes. Radiologists have
long relied on their own judgment and have read these medical images through their own eyes, which is
often subjective, time-consuming, and inter-observer variability is also likely to occur. Applications built on
artificial intelligence (AI), or more specifically, deep learning (DL)-based algorithms, have radically changed
the medical imaging field over the last couple of years and could potentially be used to automate the diagnosis
process, offering prompt, trustworthy, and unbiased assessments. Despite such developments, most existing
systems that rely on AI are constrained, especially when it comes to classification accuracy and robustness
across different datasets. To overcome these problems, the article in this chapter presents a more effective
DL model with a specifically designed architecture that aims to improve the classification of brain tumors.
The specified methodology is based on preprocessing and data normalization steps that reduce noise and
level out the data intensity, enabling effective feature extraction from the MRI images. This will increase
the accuracy of the later classification. The primary component of the proposed methodology is an adapted
version of DenseNet-201, designed explicitly for the four-class brain tumor classification. To achieve optimal
performance, the conventional output layer of DenseNet-201 was replaced with a Global Average Pooling
(GAP) layer, designed to address the issues of vanishing gradients and overfitting commonly encountered
during the training of deep networks. The architectural adjustment helps to combine the features and increase
the overall generalization capacity of the model. The model was thoroughly tested using two datasets: one
publicly available dataset on Figshare and a locally available dataset comprising a total of 3,504 T1-weighted
contrast-enhanced MRI (T1-w MRI) images. The results of the experiment provided the proposed model
with a general accuracy of 100 percent, which was higher than that of the existing comparative methods.
Such results support the idea that complex architectural adjustments with the broader preprocessing strategy
can be effective, and why deep neural networks can be viewed as trustworthy diagnostic tools in clinical
neuro-oncology, potentially achieving extremely high accuracy.
Keywords: Brain tumor; Deep Learning; DenseNet201; Multi Classification; Transfer Learning