Page 101 - AN-4-4

Advanced Neurology Diffusion model for brain tumor classification

the model benefits significantly from the high-quality same pre-processed dataset for consistent evaluation. The
synthetic images generated using DDM. The stability of comparison focused on key metrics, including accuracy,
both sets of curves underscores the effectiveness of the sensitivity, specificity, and F1 score. The performance of
DDM-based data augmentation strategy in enhancing each model is summarized in Table 4.
model generalization and convergence. The learning
progression reflects that the CDCNN model, when trained 4. Discussion
on a combination of real and synthetic datasets, efficiently The proposed CDCNN model, when combined with DDM
captures the discriminative features of brain tumor classes. for synthetic data augmentation, achieved a remarkable
3.4. Performance metrics and receiver operating classification accuracy of 96.2%, outperforming other
characteristic analysis state-of-the-art architectures and prior works in the
field. This is a substantial improvement over the work by
In addition to accuracy, sensitivity, specificity, F1 score,
and model convergence/learning curves, the positive
predictive value (PPV) and negative predictive value
(NPV) were computed for each tumor class, as shown in
Table 3. The CDCNN model achieved high PPV and NPV
values across all tumor types, indicating a strong ability to
correctly identify both positive and negative cases.
The AUC values ranged from 0.962 to 0.991,
underscoring the excellent discriminatory power of
the model. Figure 7 displays the receiver operating
characteristic curves for each tumor class, showing high
true positive rates and low false positive rates, consistent Figure 6. Training and validation loss curves illustrating stable and
decreasing trends, indicating effective model optimization
with the high sensitivity and specificity metrics.
3.5. Hexagonal radar chart visualization
To provide a holistic view of the model’s performance
across all key metrics—accuracy, sensitivity, specificity, F1
score, PPV, and NPV—a hexagonal radar chart was plotted
(Figure 8). This visualization allows for direct comparison
of the balance between sensitivity and specificity, as
well as the trade-off between precision and recall. The
near-uniform shape and large coverage area in the chart
demonstrate the consistent and balanced performance of
the CDCNN model across metrics, compared to baseline
models that show relatively irregular performance profiles.
3.6. Comparative analysis of the CDCNN model with
standard models
To evaluate the performance of the CDCNN model, a
comparative analysis was conducted against established Figure 7. Receiver operating characteristic curves for brain tumor
classification across the four tumor classes: Glioma, meningioma,
architectures: ResNet50, VGG16, VGG19, and pituitary tumor, and no tumor. The areas under the curves represent 98%
InceptionV3. These models were trained and tested on the confidence intervals.

Table 3. Expanded performance metrics on synthetic datasets

Class Precision (PPV [%]) Recall (sensitivity [%]) F1 score (%) Specificity (%) NPV (%) AUC
Glioma 89.0 91.0 90.0 96.3 95.8 0.964
Meningioma 92.0 89.0 91.0 96.9 95.5 0.962
Pituitary tumor 88.0 95.0 91.0 96.5 97.2 0.973
No tumor 98.0 97.0 97.0 98.5 97.8 0.991
Abbreviations: AUC: Area under the curve; NPV: Negative predictive value; PPV: Positive predictive value.

Volume 4 Issue 4 (2025) 95 doi: 10.36922/AN025130025