Skip to content

Latest commit

 

History

History
87 lines (59 loc) · 6.08 KB

TECHNICAL_REPORT.md

File metadata and controls

87 lines (59 loc) · 6.08 KB

Explainable Deep Learning for Multi-Class Brain Tumor Classification: A Comparative Study of Visual Interpretability

Date: January 2026 Repository: Explainable-Brain-MRI

Abstract

Deep learning has achieved dermatologist-level accuracy in medical image classification. However, the lack of transparency in "black-box" models hinders their adoption in clinical workflows, where understanding the rationale behind a diagnosis is as critical as the diagnosis itself. This report presents a robust pipeline for multi-class brain tumor classification (Glioma, Meningioma, Pituitary, No Tumor) using T1-weighted MRI scans. We integrate Grad-CAM (Gradient-weighted Class Activation Mapping) to generate visual explanations, allowing for the validation of model focus against radiological ground truth. Our approach aims to bridge the gap between high-performance ML and clinical trust, aligning with industry standards for transparent AI in healthcare.

1. Introduction

1.1 The Clinical Challenge

Brain tumors, including Gliomas and Meningiomas, require rapid and accurate triage. Structural MRI, particularly T1-weighted imaging, is the standard modality for assessing anatomical boundaries and classifying tissue. While manual interpretation is time-consuming and subject to inter-observer variability, automated systems must provide more than just a probability score to be clinically useful.

1.2 The "Black Box" Problem

Convolutional Neural Networks (CNNs) are powerful but opaque. In high-stakes domains like neuro-oncology, a false positive driven by image artifacts (e.g., a skull label or scanner noise) can have severe consequences. Explainable AI (XAI) addresses this by visualizing the features driving the model's predictions.

2. Related Work

Recent literature highlights the converging trend of high-accuracy CNNs and post-hoc explainability:

  • Iftikhar et al. (2025) demonstrated that while CNNs can reach 99% accuracy in tumor detection, the integration of SHAP and Grad-CAM is essential for validating that the model is detecting the tumor and not background noise.
  • Gharaibeh et al. (2025) utilized Xception-based netwoks with Grad-CAM to differentiate between tumor subtypes, emphasizing the need for region-based visual explanations.
  • Islam et al. (2025) proposed Grad-CAM++ improvements for better localization of lesion boundaries, effectively "segmenting" the tumor without explicit segmentation labels.

Our work builds on these foundations, providing a streamlined, reproducible framework for applying these techniques to standard T1-weighted MRI datasets.

3. Methodology

3.1 Dataset

We utilize a widely recognized public dataset comprising MRI scans categorized into four classes:

  1. Glioma: Glial cell tumors, often with irregular boundaries.
  2. Meningioma: Typically well-circumscribed dural-based tumors.
  3. Pituitary: Tumors in the sellar region.
  4. No Tumor: Healthy brain tissue.

Preprocessing: Images are resized to 224x224, normalized using ImageNet statistics (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), and augmented with random rotations and flips to improve generalization.

3.2 Model Architecture: ResNet18

We employ ResNet18 with transfer learning. The choice of ResNet18 is deliberate:

  • Efficiency: It offers a lightweight architecture suitable for rapid inference.
  • Residual Learning: Mitigates the vanishing gradient problem, allowing for effective feature extraction from complex medical images.
  • Transfer Learning: Pre-training on ImageNet provides robust low-level feature detectors (edges, textures) which are then fine-tuned for high-level MRI features.

3.3 Explainability Module: Grad-CAM

We implement Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize saliency. Mathematically, for a given class $c$ and feature map activation $A^k$ in the final convolutional layer:

  1. We compute the neuron importance weights $\alpha_k^c$ by global average pooling the gradients of the score $y^c$ with respect to feature maps $A^k$: $$ \alpha_k^c = \frac{1}{Z} \sum_i \sum_j \frac{\partial y^c}{\partial A_{ij}^k} $$
  2. We compute the weighted combination of forward activation maps, followed by a ReLU to keep only features that have a positive influence on the class of interest: $$ L_{Grad-CAM}^c = ReLU\left(\sum_k \alpha_k^c A^k\right) $$

This results in a coarse heatmap of the same size as the convolutional feature maps (e.g., 7x7), which is then upsampled to the input image resolution (224x224) for overlay.

4. Discussion & Future Directions

4.1 Interpretation of Saliency Maps

By overlaying Grad-CAM heatmaps on T1-weighted images, we can verify "Clinical Validity":

  • Success Case: The heatmap lights up the hyperintense tumor region.
  • Failure Case: The heatmap focuses on the skull, eyes, or text annotations on the scan. This verification step is crucial for deploying models in real-world settings where "Right for the wrong reasons" is unacceptable.

4.2 Industry Relevance

For pharmaceutical and medical device companies, this pipeline demonstrates:

  1. Regulatory Readiness: XAI is increasingly requested by regulatory bodies (FDA, EMA) for AI-based Medical Drives (SaMD).
  2. Workflow Integration: The ability to present a radiologist with a "second opinion" supported by visual evidence.

4.3 Future Work

  • Quantitative XAI Evaluation: Measuring the IoU (Intersection over Union) between the Grad-CAM heatmap and manual tumor segmentation masks.
  • Perturbation Methods: Integrating LIME (Local Interpretable Model-agnostic Explanations) to test model robustness against noise.
  • Multimodal Fusion: Combining T1 with T2/FLAIR images for richer input data.

5. Conclusion

This project establishes a baseline for Explainable Brain MRI Classification. By prioritizing interpretability alongside accuracy, we provide a framework that is not only valid from a machine learning perspective but also meaningful and trustworthy for clinical stakeholders.