A comparative study evaluating classical machine learning models (Logistic Regression, Random Forest, XGBoost) and a fully connected Artificial Neural Network (ANN) for multi-class brain tumor classification on MRI scans.
This project classifies magnetic resonance imaging (MRI) scans into four distinct classes:
- Glioma
- Meningioma
- Pituitary Tumor
- No Tumor
The primary objective was to establish baseline performance metrics using classical ML algorithms on flattened spatial features, evaluate a fully connected ANN, and identify the strengths, failure modes, and spatial limitations of vector-based computer vision pipelines in medical diagnostics.
All models were evaluated on an unseen, balanced test dataset of 1,600 images (400 images per class) preprocessed to
| Model | Test Accuracy | Macro Precision | Macro Recall | Macro F1-Score | Macro ROC-AUC | Glioma Recall |
|---|---|---|---|---|---|---|
| Random Forest | 85.81% | 86.92% | 85.81% | 0.8538 | 0.9620 | 64.00% |
| XGBoost | 85.56% | 85.20% | 85.56% | 0.8514 | 0.9639 | 69.00% |
| ANN (Dense Network) | 84.50% | 84.25% | 84.50% | 0.8429 | 0.9521 | 70.25% |
| Logistic Regression | 78.63% | 78.75% | 78.63% | 0.7838 | 0.9192 | 61.75% |
-
Aspect Ratio Preservation: Resizing non-square MRI scans directly into fixed square dimensions (
$224 \times 224$ ) warps spatial tumor geometry. Implementing letterbox padding (padding borders with zero-value pixels prior to resizing) maintains the structural integrity of brain tissues and tumor boundaries. -
Normalization & Gradient Flow: Unscaled pixel values (
$0$ –$255$) fed into dense neural layers cause immediate gradient instability. Verifying data types (uint8vsfloat32) prevents critical failures such as double normalization (which squishes pixel ranges to$0.0$ –$0.0039$, causing vanishing gradients) or zero-gradient saturation. -
Label Order Consistency: High validation performance (~83.5%) paired with near-random test accuracy (~25.0%) often points to a label permutation error rather than model overfitting. Ensuring index-to-class mappings (
0: glioma,1: meningioma,2: no_tumor,3: pituitary) remain identical acrosstf.data.Datasetgenerators and evaluation scripts is essential.
-
Curse of Dimensionality: Flattening a
$224 \times 224$ image yields$50,176$ continuous features per sample. Traditional models (especially tree-based ensembles) encounter severe memory bottlenecks during tree-split calculations across large feature spaces. -
Spatial Blindness: Vectorizing 2D image matrices destroys pixel locality—the physical relationship between adjacent pixels is lost. As a result, classical models perform well on well-defined structural boundaries (
pituitaryandno_tumor$>90%$ F1-score), but struggle with diffuse, irregular tumor margins (gliomarecall drops to$64.0%$ ). - Performance Clustering: Non-linear models using 1D flattened vectors hit a natural performance ceiling at 84.5%–85.8% accuracy, demonstrating the upper boundary of non-convolutional approaches on raw pixel arrays.
-
Tree Ensembles vs. Neural Networks: Random Forest achieved the highest overall accuracy (85.81%), but suffered from significant false negatives on aggressive tumors (missing
$36%$ of gliomas). -
Glioma Sensitivity in ANNs: The fully connected ANN achieved the highest Glioma Recall (70.25%) among all models. Non-linear activation combinations (
ReLU+BatchNormalization) allowed the neural network to learn complex intensity thresholds better than orthogonal decision tree splits.