adwitiyashukla/hybrid-dl-multiclass-brain-tumor-classification-mri

Hybrid DL for Multi-Class Brain Tumor Classification using MRI: Healthcare Management System

Upload an axial brain MRI slice and the system runs a six stage classical image processing pipeline, classifies the slice into one of four classes, and returns a Grad-CAM++ attribution map alongside the prediction. Results can be saved against patient records, reviewed by a clinician, and exported as a report.

Source code: https://github.com/adwitiyashukla/hybrid-dl-multiclass-brain-tumor-classification-mri

Classes

Glioma, meningioma, pituitary adenoma, no tumor.

Processing pipeline

1. 1.Skull stripping by Otsu thresholding and morphological cleanup
2. 2.Bias field correction by homomorphic low pass division
3. 3.Non-local means denoising for Rician noise
4. 4.Intensity normalisation by z-score inside the brain mask
5. 5.CLAHE local contrast enhancement
6. 6.Midsagittal alignment and bilateral asymmetry mapping

The asymmetry map exploits the approximate bilateral symmetry of the healthy brain. The slice is aligned to its own midsagittal plane, mirrored, and subtracted, so symmetric healthy tissue cancels and pathology remains.

Model

EfficientNet-B0 over three channels (processed slice, brain mask, asymmetry map), with a CBAM attention block and a parallel branch encoding 43 handcrafted features. A learned gate mixes the two streams and its value is displayed with each prediction, indicating how much the result relied on learned versus handcrafted evidence.

Results

Measured on a leak free test split of 1180 images, with bootstrap confidence intervals over 2000 resamples.

Per class F1: glioma 0.901, meningioma 0.909, no tumor 0.891, pituitary 0.974.

Why leak free: roughly 26 percent of the supplied test split turned out to be duplicates of training images, concentrated heavily in the no tumor class at 77.5 percent. Those images were removed before scoring. Evaluating on the full supplied split instead gives macro F1 0.9412, which is optimistic. Details, method and the reproduction script are in the GitHub repository.

Glioma is the weakest class at 0.824 recall. It is the target for future work.

Tabs

• —Dashboard: throughput, class distribution and review status
• —New scan: upload, pipeline visualisation, prediction, attribution map
• —Patients: patient list, scan history, clinician review and override
• —Reports: exportable report per scan
• —About: method summary and limitations

Important

Research prototype. Not a medical device. Not clinically validated. Predictions must not be used for diagnosis or treatment. Do not enter real patient identifiers.