Page 102 - AN-4-4

Advanced Neurology Diffusion model for brain tumor classification

iterative refinement ensures training stability, avoids mode
collapse, and produces highly detailed synthetic images
that preserve tumor morphology and anatomical features.
Moreover, DDM does not require adversarial competition
between generator and discriminator networks, which is
the primary source of instability in GAN training. Instead,
it uses a maximum likelihood-based training objective,
resulting in a more predictable and stable convergence.
This stability ensures that the generated synthetic MRI
slices maintain consistent quality across large batches,
enabling the downstream CDCNN classifier to learn from
a broader and more representative feature space.
4.1. Rationale for combining the DDM with a CDCNN
The CDCNN architecture is inherently powerful for medical
imaging tasks because it integrates deep convolutional
Figure 8. Hexagonal radar chart of model performance illustrating the feature extraction with conditional processing, allowing
balanced performance of the conditional deep convolutional neural it to focus on class-specific discriminative features. When
network model across accuracy, sensitivity, specificity, F1 score, positive trained on DDM-generated images alongside real MRI
predictive value (PPV), and negative predictive value (NPV). Larger data, the CDCNN benefits from:
enclosed areas indicate superior performance.
(i) Increased dataset size: DDM produced 10,000
synthetic MRI images across four tumor classes,
Table 4. Summary of different models’ results effectively expanding the dataset and improving the

Model Accuracy Sensitivity Specificity F1 Score network’s generalization ability.
(%) (%) (%) (%) (ii) Improved feature diversity: DDM-generated images
CDCNN 96.2 95.8 96.7 96.0 introduce subtle variations in tumor shapes, textures,
ResNet50 91.5 90.8 92.0 91.2 and intensities, enabling CDCNN to learn richer
decision boundaries.
VGG16 89.7 88.9 90.5 89.3 (iii) Balanced class representation: Synthetic augmentation
VGG19 90.3 89.7 91.0 90.0 corrected class imbalances in the original dataset,
InceptionV3 92.8 91.9 93.5 92.2 reducing bias and improving per-class recall and
Abbreviations: CDCNN: Conditional deep convolutional neural precision.
network; VGG: Visual Geometry Group.
This synergy between high-fidelity synthetic data and a
robust DL architecture resulted in balanced performance
Onakpojeruo et al., where the same CDCNN model was across all tumor classes, with particularly strong results
52
used with Pix2Pix for synthetic data generation, achieving for the challenging glioma class, which historically suffers
an accuracy of only 86%. The 10% increase in accuracy from high misclassification rates.
strongly highlights the effectiveness of DDM in generating
high-quality, realistic, and diverse synthetic datasets that The DDM-CDCNN framework demonstrated high
significantly boost classification performance. sensitivity and specificity across all four tumor categories.
For example, the no tumor class achieved a specificity of
Traditional GAN-based methods, while powerful, are 97%, indicating the model’s ability to correctly identify
known to suffer from issues, such as mode collapse, unstable negative cases and avoid false positives. The balanced F1
training dynamics, and sensitivity to hyperparameter scores further underscore the robustness of the model,
tuning. These limitations often result in synthetic images ensuring reliable classification without sacrificing either
that either lack diversity or fail to capture fine structural precision or recall. Importantly, the ablation study revealed
details crucial for accurate medical image classification. In that removing the DDM-generated synthetic data reduced
contrast, DDM operates using a progressive noise addition accuracy by 8.5%, confirming its critical role in enhancing
and removal process. During the forward diffusion process, classification performance. Statistical tests further
Gaussian noise is added to the image over several timesteps, validated these results: Paired t-tests and ANOVA showed
gradually destroying image structure. The reverse diffusion significant differences (p<0.05) between the proposed
process then reconstructs the image by denoising step- method and baseline models, while confidence intervals
by-step, conditioned on a learned distribution. This demonstrated the stability of the performance metrics.

Volume 4 Issue 4 (2025) 96 doi: 10.36922/AN025130025