Fuzzy deep neural network for brain tumor classification
Lead researcher, with Dr. Devinder Kaur
95.8% overall accuracy across four classes, externally validated through the University of Pennsylvania.
95.8%overall accuracy
The problem
Tumour margins on MRI are soft and blurred. Hard clustering forces a boundary where the image does not contain one, and the segmentation error carries straight into whatever classifier consumes it.
The data
MRI brain scans across four classes: glioma, meningioma, pituitary and no tumor.
Preprocessing covered noise reduction, intensity normalisation and registration.
Approach
Fuzzy C-Means clustering to segment tumour regions, which assigns partial membership rather than a hard label and so handles an ambiguous boundary on its own terms.
The segmentation path is UNet into a fuzzy layer into UNet++: the first pass proposes a boundary, the fuzzy layer expresses how certain each pixel’s membership is, and UNet++ refines against that rather than against a hard mask. Statistical information and loss values feed a knowledge base that the inference path queries.
A CNN for classification — convolutional and max-pooling stages, a dense 1024 layer, transfer learning — with hybrid integration at both feature and decision level, followed by clinical validation.
Results
95.8% overall accuracy, a 7.3% improvement over traditional methods. The hybrid model reached 97.87% at peak — that figure is a peak, not the headline.
The model-level ROC comparison is the stronger evidence, because it puts the hybrid against what it replaces rather than reporting it alone:
| Method | AUC |
|---|---|
| Proposed hybrid | 0.926 |
| CNN only | 0.871 |
| Traditional ML | 0.815 |
Converged within 25 epochs.
| Class | Accuracy | Precision | Recall | F1 |
|---|---|---|---|---|
| Glioma | 96.2% | 95.8% | 95.3% | 95.5% |
| Meningioma | 94.8% | 94.2% | 93.9% | 94.0% |
| Pituitary | 95.5% | 95.1% | 94.8% | 94.9% |
| No tumor | 97.2% | 96.8% | 96.5% | 96.6% |
Externally validated through the University of Pennsylvania.
Limitations and what I'd do differently
Accuracy is reported per class on a held-out split; the dataset size and class balance are not published here, and without them the per-class figures should be read as indicative rather than definitive.
Meningioma is the weakest class at 94.8%, and it is the one whose margins are most often ambiguous. The error concentrates exactly where the method was supposed to help most, which is the result worth investigating next.
Clinical validation here means agreement with expert labels on the available data, not a prospective study. Nothing about deployment in a clinical workflow is demonstrated.
Artifacts
Figures from the paper are the publisher’s and are not reproduced here.