The diagnosis of brain tumor from MRI is a complex problem considered to be one of the most important aspects in medical diagnosis since timely detection along with correct categorization can considerably enhance patients’ prognosis. This paper aims at investigating the use of the two main deep learning architectures, namely ResNet and DenseNet, for brain tumor diagnosis. Specifically, these pre-trained models were used through transfer learning and fine-tuned on brain tumor MRI scans dataset to evaluate the performance metrics such as accuracy, sensitivity, specificity, precision, and F1-score. ResNet fixes the vanishing gradient problem that is inherent to deep learning and lets one train deeper models. DenseNet as a result of the dense connectivity between the layers; enhances feature reuse and gradient flow, it therefore makes better use of the parameters. The experimental results show that DenseNet achieved a higher validation accuracy of $95 \%$, while ResNet achieved $85 \%$ even with a seemingly more difficult distribution, with only $12 \%$ and $37 \%$, pointing to better generalization and stability of the receiving function. Model improving techniques included random rotations of images, flipping, and random changes in image color intensity. Precise learning rate management strategies such as learning rate scheduling helped to stabilize the training process, DenseNet among them. The confusion matrix analysis proceeding it further affirmed DenseNet’s performance in ‘No Tumor’ and ‘Tumor’ differentiation. In this objective comparison, DenseNet’s ability for better and accurate brain tumor detection when applied in a clinical environment has been postulated. Further works should consider incorporating tumor localization to the classification technique, investigate new network architectures further and, test such algorithms in authentic clinic setting to build a more efficient diagnostic tool to improve patients’ treatment.
Deep Learning-based MRI Analysis for Brain Tumor Detection: Insights from ResNet and DenseNet Models
06.11.2024
451145 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch