PruDensNet: a parameter efficient depthwise separable CNN for MRI-based brain tumor classification
-
By
-
Mithila Arman
-
Ahnaf Samin
-
A. K. M. Muzahidul Islam
-
Md Maruf Rusafi Arnob
-
Md Jahirul Islam
-
Ishtiak Al Mamoon
-
May 20, 2026
-
Clinical Scorecard: PruDensNet: An Efficient Depthwise Separable CNN for Classifying Brain Tumors Using MRI Data
At a Glance
| Category | Detail |
|---|
| Condition | Brain Tumors |
| Key Mechanisms | Depthwise-separable convolutions, channel and spatial attention, GELU activations. |
| Target Population | Patients undergoing MRI for brain tumor assessment. |
| Care Setting | Cost-latency-sensitive clinical deployments, including CPU-only radiology workstations and low-resource edge installations. |
Key Highlights
- Achieves 96.05% test accuracy and 97.27% validation accuracy on a four-class brain tumor benchmark.
- Utilizes a lightweight architecture with approximately 1.46 million parameters.
- Incorporates a robust training methodology with curriculum-regularization and data curation.
- Demonstrates a favorable accuracy-footprint trade-off for clinical workflows.
- Ensures reproducibility through standardized protocols and fixed seeds.
Guideline-Based Recommendations
Diagnosis
- Utilize MRI as the primary modality for non-invasive brain tumor assessment.
Management
- Implement PruDensNet for automated classification of brain tumors in clinical settings.
Monitoring & Follow-up
- Regularly evaluate model performance against external validation datasets.
Risks
- Be aware of potential data leakage and ensure rigorous data curation to maintain metric integrity.
Patient & Prescribing Data
Patients with suspected brain tumors requiring MRI analysis.
PruDensNet can enhance diagnostic accuracy while minimizing computational costs.
Clinical Best Practices
- Adopt a standardized data curation pipeline to prevent duplicates and ensure label consistency.
- Utilize advanced training techniques such as MixUp and CutMix to improve model robustness.
- Implement mixed-precision training to optimize resource usage without sacrificing performance.
Related Resources & Content
