Medical Image Abnormality Classification & Spatial Localization System (MedImgSys)

📌 Clinical Context & Project Overview

Magnetic Resonance Imaging (MRI) serves as the primary diagnostic modality for intracranial neoplasm evaluation in clinical neuro-oncology. At regional healthcare institutions such as the Teaching Hospital Peradeniya, conventional radiological workflows encounter three persistent operational challenges:

1. Diagnostic Latency & Manual Reporting: High patient influx requires radiologists to manually examine multi-slice radiographic scans and hand-transcribe findings into Diagnostic Imaging Reports (DIR), inducing critical treatment delays.
2. Inter-Observer Morphological Ambiguity: Infiltrative intra-axial gliomas, extra-axial meningiomas, and intrasellar pituitary adenomas often present overlapping radiological signatures on standard T1-weighted sequences, causing diagnostic variance across observers.
3. Clinical Trust & Explainability Gap: Black-box neural networks deliver high statistical performance but fail to provide spatial provenance, hindering clinical adoption by radiologists, neuro-oncologists, and neurosurgeons.

MedImgSys resolves these challenges by introducing an end-to-end computer vision and deep learning system engineered specifically for clinical neuro-imaging workflows. The platform couples a specialized 4-stage soft-tissue preprocessing engine with dual convolutional neural backbones and Grad-CAM visual localization, delivering verifiable spatial interpretability without demanding pixel-level manual masks.

🌟 Key Performance Highlights

• 95.69% Test Accuracy: Verified across a strictly balanced evaluation set of 1,600 unseen MRI scans.
• +22.50% Preprocessing Gain: Rigorous 3-stage ablation demonstrates that custom medical preprocessing elevates baseline accuracy from 73.19% (pure raw) to 95.69% (full pipeline).
• Architectural Parity with 78.5% Fewer FLOPs: Benchmarks ResNet-18 against EfficientNet-B0, showing compound scaling maintains diagnostic accuracy (95.62% vs 95.69%) while trimming parameter footprint by 64.1% and computational complexity by 78.5%.
• Zero-Annotation Spatial Localization: Generates high-fidelity gradient-weighted activation heatmaps that clinically coincide with anatomical tumor lesions.

🏗 System Architecture

The MedImgSys framework operates as a modular, feed-forward pipeline from raw patient acquisition to explainable diagnostic reporting:

Pipeline Stages

1. Acquisition & Skull Stripping: Raw heterogeneously sized DICOM/PNG MRI scans pass through automated Otsu contour thresholding to excise non-brain background margins and skull tissue.
2. Soft-Tissue Conditioning: Bilateral edge-preserving filtering eliminates magnetic field high-frequency noise while LAB-space CLAHE maximizes soft-tissue lesion contrast.
3. Stochastic Augmentation: Geometric and photometric transforms (Albumentations) simulate scanner field variances and patient alignment differences.
4. Dual Backbone Classification: ResNet-18 (residual baseline) and EfficientNet-B0 (compound scaling) predict 4 diagnostic categories with class-weighted cross-entropy loss.
5. Grad-CAM Saliency Localization: Backpropagated feature gradients extract spatial attention maps, overlaying color-coded heatmaps on original anatomical scans for clinical auditability.

📊 Dataset Specifications

The system is developed and benchmarked on the Kaggle Brain Tumor MRI Dataset (masoudnickparvar/brain-tumor-mri-dataset), comprising 7,022 high-resolution T1-weighted contrast-enhanced MRI scans.

Diagnostic Class Profiles

Partitioning & Stratification

To prevent data contamination and guarantee unbiased evaluation, scans are partitioned using a deterministic global seed (42):

• Official Benchmark Test Set: 1,600 scans (Strictly balanced: 400 images per class, 25.0% each).
• Training Set: 4,480 scans (Stratified 80% split from training pool: 1,320 Glioma, 1,339 Meningioma, 1,595 No Tumor, 1,456 Pituitary).
• Validation Set: 1,120 scans (Stratified 20% split from training pool: 330 Glioma, 335 Meningioma, 399 No Tumor, 364 Pituitary).
• Class Balancing: To counterbalance class frequency variations during optimization, dynamic inverse-frequency weighting is applied:

Weight_c = N_total / (C × N_c)
(where N_total is the total training dataset count, C = 4 is the number of classes, and N_c is the sample count for class c)

🔬 Medical Image Preprocessing Engine

Raw clinical MRI acquisitions routinely exhibit non-standardized black borders, magnetic coil inhomogeneities, and subtle soft-tissue contrast gradients. The MedImgSys 4-Stage OpenCV Preprocessing Engine conditions raw inputs prior to tensor ingestion:

Algorithmic Breakdown

1. Contour-Based Skull Stripping (crop_brain_contour):
   • Grayscale conversion and Gaussian smoothing (5 × 5, σ = 0) to eliminate salt-and-pepper noise.
   • Binary thresholding via Otsu’s method combined with morphological erosion and dilation (2 iterations, 3 × 3 kernel).
   • External contour extraction to identify the largest cranial tissue boundary.
   • Tight bounding box crop that discards 40%–60% of uninformative dark background pixels.
2. Edge-Preserving Bilateral Denoising:
   • Applied with diameter d = 9, σ_color = 75, σ_space = 75.
   • Replaces pixel values via a bilateral Gaussian weight combining spatial proximity and radiometric photometric distance, effectively smoothing acquisition grain without degrading sharp tumor-brain margins.
3. CIE LAB Contrast Limited Adaptive Histogram Equalization (CLAHE):
   • Translates images into the perceptual CIE LAB color space.
   • Confines adaptive equalization exclusively to the Luminance (L*) channel with clipLimit = 2.0 and tileGridSize = (8, 8) to prevent noise amplification in uniform brain tissue.
   • Re-merges with chromaticity channels (a*, b*) and converts back to RGB.
4. Standardization & ImageNet Normalization:
   • Resizes via area interpolation (cv2.INTER_AREA) to a standardized 224 × 224 × 3 tensor.
   • Standardizes channel distributions according to ImageNet statistics (mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225]).

🔄 Data Augmentation Strategy

To simulate variable clinical imaging protocols and patient positioning inside magnetic resonance bores, batches undergo stochastic online augmentations via Albumentations:

🧠 Deep Learning Architectures

MedImgSys evaluates two distinct neural network paradigms to establish the trade-off between absolute diagnostic throughput and embedded edge deployability:

1. ResNet-18 (Residual Learning Baseline)

• Design: 18-layer residual architecture using identity skip connections:

  y = F(x, {W_i}) + x

• Advantage: Guarantees direct gradient propagation during backpropagation, circumventing vanishing gradient problems.

• Complexity: 11.18 Million parameters

  1.81 GFLOPs.

2. EfficientNet-B0 (Compound Scaling Modern Benchmark)

• Design: Utilizes inverted residual bottleneck blocks (MBConv) with depthwise separable convolutions and Squeeze-and-Excitation (SE) channel-attention modules.
• Compound Scaling: Systematically balances network depth, channel width, and input resolution:

  depth: d = α^φ, width: w = β^φ, resolution: r = γ^φ
  (subject to α · β² · γ² ≈ 2)

• Complexity: 4.01 Million parameters (64.1% reduction)

  0.39 GFLOPs (78.5% reduction).

Training Configuration

• Optimizer: AdamW (Initial learning rate = 3 × 10⁻⁴, weight decay = 1 × 10⁻⁴)
• Learning Rate Schedule: CosineAnnealingLR (T_max = 10 epochs, min_lr = 1 × 10⁻⁶)
• Precision: Automatic Mixed Precision (AMP) via PyTorch autocast and GradScaler
• Hardware: NVIDIA RTX 6000 Ada Generation (51.5 GB VRAM)

• Batch Size: 32

  Epochs: 10

📈 Quantitative Model Benchmarking

Both architectures were subjected to rigorous evaluation against the 1,600 unseen scans of the official test partition:

Performance Summary Table

Clinical & Architectural Insights

1. Statistical Accuracy Parity: EfficientNet-B0 matches ResNet-18 within 0.07% test accuracy (95.62% vs 95.69%) and 0.0008 macro F1-score (0.9555 vs 0.9563).
2. Computational Footprint Reduction: EfficientNet-B0 eliminates 64.1% of parameter weight storage (16.4 MB vs 44.8 MB) and cuts computational operations by 78.5% (0.39 vs 1.81 GFLOPs).
3. Deployment Strategy:
   • Centralized Hospital PACS Server: ResNet-18 yields minimal batch latency (5.04 ms/image) for high-throughput radiology processing queues.
   • Point-of-Care & Mobile MRI Consoles: EfficientNet-B0 provides an optimal footprint for resource-constrained edge workstations, tablets, and embedded hardware.

🧪 3-Stage Ablation Study

To mathematically decouple the contribution of the medical preprocessing engine from the benefits of data augmentation, a controlled 3-stage ablation study was conducted using ResNet-18 under identical optimization protocols:

Experimental Results

Ablation Findings

• The Invariance Gap (+20.12%): Training unaugmented raw scans results in severe network memorization of scanner noise and background pixels, capping test accuracy at 73.19%. Injecting stochastic transforms forces the network to learn invariant geometrical features, boosting accuracy to 93.31%.
• The Medical Contrast Boost (+2.38% / +22.50% total): Combining skull contour excision, bilateral filtering, and LAB CLAHE drives test accuracy to 95.69%. By stripping irrelevant skull/background pixels and accentuating subtle parenchymal boundary gradients, the network reliably differentiates difficult borderline cases.

🔍 Spatial Localization & Explainability (Grad-CAM)

To provide verifiable spatial interpretability without manual pixel-level segmentation masks, MedImgSys implements Gradient-Weighted Class Activation Mapping (Grad-CAM):

Mathematical Formulation & Algorithmic Steps

1. Target Feature Map Gradients:
   Compute the gradient of the predicted class score y^c with respect to feature activation maps A^k of the final convolutional layer (resnet18.layer4[-1] or effnet.features[-1]):

   Gradient = ∂(y^c) / ∂(A^k)

2. Neuron Importance Weights (α_k^c):
   Apply Global Average Pooling (GAP) across spatial height U and width V:

   α_k^c = (1 / (U × V)) × ∑_i ∑_j [ ∂(y^c) / ∂(A_i,j^k) ]

3. Linear Combination and Rectification (ReLU):
   Compute the importance-weighted sum of forward activation maps, passing through a ReLU activation to preserve features that contribute positively to the target class:

   L_Grad-CAM^c = ReLU( ∑_k α_k^c · A^k )

4. Heatmap Normalization & Color Blending:
   Normalize the activation map to [0, 1], upsample to 224 × 224, apply OpenCV COLORMAP_JET, and blend with the preprocessed anatomical MRI scan:

   I_overlay = 0.4 × I_heatmap + 0.6 × I_anatomical

Clinical Validation Across Classes

• Glioma: Activation maps localize deeply within intra-axial white matter parenchyma, following the irregular, infiltrative boundaries characteristic of glial neoplasms.
• Meningioma: Activation concentrates along extra-axial dural interfaces and peripheral convexities, pinpointing the characteristic dural tail enhancement.
• Pituitary Tumor: Heatmaps focus precisely on the sellar and suprasellar regions at the skull base, correctly targeting adenomas around the pituitary fossa.
• No Tumor: Visualizations display low-magnitude, diffuse, non-focal activations across normal cerebral parenchyma, confirming the absence of pathological focal lesions.

📂 Repository Organization

e22-co543-Medical-Image-Abnormality-Cassification-Segmentation/
├── README.md                                          # Master repository documentation
├── code/
│   ├── README.md                                      # Codebase manual & execution guide
│   ├── Updated notebook with comparison/
│   │   ├── Brain_Tumor_MRI_Classification_and_Explainability (1).ipynb  # Master notebook
│   │   ├── best_resnet18 (1).pth                      # ResNet-18 weights (95.69% Test Acc)
│   │   ├── best_efficientnet_b0 (1).pth               # EfficientNet-B0 weights (95.62% Test Acc)
│   │   ├── best_resnet18_pureraw.pth                  # Ablation Stage 1 weights (73.19%)
│   │   └── best_resnet18_raw_aug.pth                  # Ablation Stage 2 weights (93.31%)
│   ├── New model/                                     # Initial exploratory checkpoints
│   │   ├── Brain_Tumor_MRI_Classification_and_Explainability.ipynb
│   │   ├── best_resnet18.pth
│   │   ├── best_resnet18_raw.pth
│   │   └── best_efficientnet_b0.pth
│   └── preprocessing/
│       └── preprocessing-pipeline-visulaization.ipynb # Preprocessing pipeline visualizer
└── docs/                                              # GitHub Pages publication root
    ├── _config.yml                                    # Jekyll site configuration
    ├── README.md                                      # GitHub Pages site documentation (this file)
    ├── data/
    │   └── index.json                                 # Team metadata & project tags
    └── images/
        ├── banner.svg                                 # High-resolution vector header banner
        ├── system_architecture.svg                    # Full architecture pipeline diagram
        ├── preprocessing_workflow.svg                 # 4-stage preprocessing flowchart
        ├── model_benchmark_comparison.svg             # ResNet-18 vs EfficientNet-B0 benchmark
        ├── ablation_study_chart.svg                   # 3-stage ablation study comparison
        └── gradcam_workflow.svg                       # Grad-CAM localization & explainability

🚀 Installation & Reproducibility

1. Clone the Repository

git clone https://github.com/cepdnaclk/e22-co543-Medical-Image-Abnormality-Cassification-Segmentation.git
cd e22-co543-Medical-Image-Abnormality-Cassification-Segmentation

2. Set Up Virtual Environment

# Create and activate environment
python -m venv venv

# Windows:
.\venv\Scripts\activate
# Linux / macOS:
source venv/bin/activate

3. Install Dependencies

# PyTorch with CUDA acceleration (adjust CUDA version if needed):
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

# Core computer vision, processing, and evaluation packages:
pip install albumentations opencv-python numpy pandas matplotlib seaborn scikit-learn jupyterlab

4. Dataset Setup & Execution

1. Download the Kaggle Brain Tumor MRI Dataset.
2. Extract the dataset into an archive/ folder in the project root:

   archive/
   ├── Training/
   │   ├── glioma/
   │   ├── meningioma/
   │   ├── notumor/
   │   └── pituitary/
   └── Testing/
       ├── glioma/
       ├── meningioma/
       ├── notumor/
       └── pituitary/
   

3. Launch Jupyter Lab:

   jupyter lab
   

4. Open and run:

   code/Updated notebook with comparison/Brain_Tumor_MRI_Classification_and_Explainability (1).ipynb
   

👥 Team Members

This project was developed by Group 14 as part of the CO5430 Computer Vision course at the Department of Computer Engineering, University of Peradeniya:

Hansara S. H. S. E/22/130 e22130@eng.pdn.ac.lk

H. M. Liyanage E/22/211 e22211@eng.pdn.ac.lk

D. M. N. N. Bandara E/22/044 e22044@eng.pdn.ac.lk

Weerasinghe W. P. T. H. E/22/421 e22421@eng.pdn.ac.lk

🔗 Project Links & Resources

• Project Repository
• GitHub Pages Web Site
• Department of Computer Engineering
• Faculty of Engineering, University of Peradeniya

📜 Key Academic References

1. Selvaraju, R. R., et al. (2017). Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. IEEE International Conference on Computer Vision (ICCV).
2. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition. IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
3. Tan, M., & Le, Q. V. (2019). EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. International Conference on Machine Learning (ICML).
4. Zuiderveld, K. (1994). Contrast Limited Adaptive Histogram Equalization. Graphics Gems IV, Academic Press Professional, Inc.
5. Tomasi, C., & Manduchi, R. (1998). Bilateral Filtering for Gray and Color Images. IEEE International Conference on Computer Vision (ICCV).
6. Nickparvar, M. (2021). Brain Tumor MRI Dataset. Kaggle Datasets.