From Explanations to Architecture: Explainability-Driven CNN Refinement for Brain Tumor Classification in MRI

This paper proposes an explainability-driven framework that utilizes Grad-CAM to iteratively refine and prune a CNN architecture for brain tumor classification, achieving high accuracy and strong generalization while ensuring model transparency and clinical trustworthiness.

The Gist

Imagine you are trying to teach a robot to spot different types of brain tumors in MRI scans. In the past, scientists built these robots (AI models) to be incredibly complex, like massive skyscrapers with thousands of floors. These "skyscrapers" were very good at getting the right answer, but they were black boxes. Even the people who built them couldn't explain why the robot made a specific decision. It was like asking a wizard, "Why did you cast this spell?" and getting a shrug in return.

The problem is that in medicine, knowing why is just as important as knowing what. If a robot says, "This patient has a tumor," a doctor needs to know, "Did it see the tumor, or did it just get confused by a shadow or a weird texture in the background?"

The Big Idea: From "Black Box" to "Glass House"

This paper introduces a new way of building these AI robots. Instead of just building a bigger, more complex skyscraper, the authors decided to build a smart, transparent house.

Here is the simple breakdown of how they did it, using some everyday analogies:

1. The Problem: The Over-Engineered Robot

Imagine you hire a team of 100 detectives to solve a crime. They all look at the evidence, but 80 of them are just staring at the wallpaper, the dust on the floor, or the color of the suspect's shoes. Only 20 are actually looking at the crime scene.

• The Old Way: Researchers kept hiring more detectives (adding more layers to the AI) to get a higher accuracy score. But because so many were looking at the wrong things (background noise), the model was "over-parameterized" (too heavy) and didn't explain its reasoning.

2. The Solution: The "Flashlight" (Grad-CAM)

The authors used a tool called Grad-CAM. Think of this as a magical flashlight that shines on the parts of the MRI scan the AI is actually looking at when it makes a decision.

• The Old Robot: When you shined the flashlight, it lit up the whole room, including the ceiling, the floor, and the tumor. It was hard to tell what mattered.
• The New Approach: They used this flashlight while the robot was learning. They asked, "Hey, why are you looking at the ceiling? That's not the tumor!"

3. The Refinement: Pruning the Tree

This is the most creative part of the paper. Usually, AI researchers build a model and then try to explain it. These authors did the opposite: They used the explanation to fix the model.

Imagine you have a tree with many branches. Some branches are heavy and full of leaves (important features), but others are dead, dry twigs that do nothing but weigh the tree down.

• The Process:
  1. They trained the robot.
  2. They used the "Flashlight" (Grad-CAM) to see which parts of the brain (and which parts of the robot's own brain) were actually useful.
  3. They found that some layers of the robot were just "dead weight"—they weren't helping identify the tumor; they were just looking at noise.
  4. They cut those layers out. They literally removed the useless parts of the AI's architecture.

4. The Result: A Leaner, Smarter Detective

By cutting out the dead branches, the robot became smaller, faster, and cheaper to run. But surprisingly, it got better at its job.

• Why? Because it was no longer distracted by the wallpaper and dust. It was forced to focus entirely on the tumor.
• The Proof: They tested this new, lean robot on two different sets of MRI scans.
  • On the first set, it got 98.2% correct.
  • On a completely new, unseen set (like a different hospital's data), it still got 94.7% correct.
  • This proves the robot learned the actual disease, not just the specific style of the first hospital's photos.

5. Double-Checking the Work

To make sure they weren't fooling themselves, they used two other "flashlights" (called SHAP and LIME) to double-check the robot's reasoning.

• SHAP is like a forensic accountant, breaking down exactly how much each pixel contributed to the final decision.
• LIME is like a local guide, checking if small changes to the image change the robot's mind in a logical way.
• All three tools agreed: The robot was looking at the tumor, not the background.

Why This Matters for You

This isn't just about math; it's about trust.

• For Doctors: They can finally see why the AI is making a diagnosis. If the AI highlights the tumor clearly, the doctor can trust it. If the AI highlights a random spot on the skull, the doctor knows to ignore it.
• For Patients: It means safer, faster, and more reliable diagnoses.
• For the Planet: Smaller models use less electricity, which is good for the environment (and fits with the UN's goal for "Good Health and Well-being").

The Takeaway

The authors showed that you don't need a giant, complicated, confusing AI to solve medical problems. Sometimes, the best way to build a smart system is to listen to its own explanations, cut out the nonsense, and let it focus on what really matters. They turned a "Black Box" into a "Glass House," making AI in healthcare safer, faster, and more trustworthy.

Technical Summary

1. Problem Statement

Current deep learning approaches for brain tumor classification in MRI scans often achieve high accuracy but suffer from two critical limitations:

• Lack of Interpretability: Deep models act as "black boxes," making it difficult for clinicians to verify if predictions are based on tumor-relevant evidence or spurious cues (e.g., background artifacts, normal tissue, or scanner noise).
• Over-Parameterization: To maximize accuracy, models are often made excessively deep and complex, increasing computational costs and hindering real-time clinical deployment.
• Passive Use of XAI: Existing Explainable AI (XAI) methods are typically used post hoc (after training) merely for visualization. They rarely influence the model's architecture or optimization process.

The authors propose a paradigm shift: using explainability not just as a visualization tool, but as an active design signal to refine and simplify the neural network architecture itself.

2. Methodology

The proposed framework follows a four-stage pipeline:

A. Data Preprocessing

Two public datasets were utilized:

1. Msoud Dataset: 7,023 grayscale MRI slices.
2. NeuroMRI Repository: 3,264 grayscale MRI slices.
   Both contain four classes: Meningioma, Glioma, Pituitary, and No-Tumor.

• Processing Steps: Images underwent cropping (to isolate the brain region via thresholding and contour detection), normalization (scaling intensity to [0, 255]), and resizing to $224 \times 224 \times 3$.

B. Baseline CNN Architecture

A standard hierarchical CNN was constructed as a baseline:

• Structure: Repeated blocks of Conv(3×3) + ReLU + MaxPool.
• Filters: Progressively increasing from 8 to 256.
• Head: Batch Normalization, Average Pooling, two Fully Connected layers (512 units each), and a Softmax output layer for 4 classes.

C. XAI-Guided Architecture Refinement (The Core Innovation)

Instead of training a complex model and then explaining it, the authors used XAI to prune the model:

1. Training: The baseline CNN is trained on the preprocessed data.
2. Relevance Quantification: Grad-CAM is applied to every convolutional block to generate heatmaps. The mean heatmap intensity is calculated for each layer to determine its "relevance score" ($\mathcal{R}_\ell$).
3. Pruning: Layers with relevance scores falling below a specific percentile threshold (the lowest contributors) are identified as "low-contribution" layers.
4. Refinement: These low-contribution layers are removed, and the network is rebuilt and retrained.

• Goal: To create a lighter, more efficient model that forces attention onto discriminative tumor regions by removing layers that rely on noise or irrelevant features.

D. Multi-Method Validation

To ensure the refined model's decisions are clinically sound, three complementary XAI methods were used:

• Grad-CAM: For spatial localization of tumor regions.
• SHAP (Shapley Additive Explanations): For attribution-based verification of feature contributions.
• LIME (Local Interpretable Model-agnostic Explanations): For instance-level verification using superpixels.

3. Key Contributions

1. Explainability-Driven Design: The paper introduces a novel workflow where XAI (specifically Grad-CAM) is integrated into the training loop to guide architectural simplification, rather than serving only as a post-training diagnostic.
2. Model Simplification without Accuracy Loss: By removing layers that contribute minimally to the prediction, the authors achieved a less complex CNN that maintains or improves classification performance.
3. Robust Generalization: The model was validated on an unseen dataset (Dataset-2) to prove it does not overfit to specific dataset artifacts.
4. Clinical Trustworthiness: The use of multiple XAI techniques (Grad-CAM, SHAP, LIME) confirms that the model focuses on anatomically correct tumor regions (e.g., extra-axial regions for meningiomas, sellar regions for pituitary tumors) rather than background noise.

4. Experimental Results

Performance Metrics:
The proposed model outperformed the initial baseline on both datasets:

Dataset	Model	Accuracy	Precision	Recall	F1-Score
Dataset-1	Initial Baseline	97.10%	97.16%	97.90%	97.93%
	Proposed (Refined)	98.21%	98.22%	98.18%	98.20%
Dataset-2 (Unseen)	Initial Baseline	92.33%	93.35%	93.83%	93.24%
	Proposed (Refined)	94.72%	94.70%	95.10%	94.63%

Qualitative Analysis:

• Grad-CAM: The refined model showed more focused attention on tumor-specific areas (e.g., intra-axial regions for gliomas, sellar regions for pituitary tumors) compared to the baseline, which often exhibited diffuse attention or focused on surrounding normal tissue.
• SHAP & LIME: Confirmed that positive contributions were linked to clinically relevant morphological features (e.g., irregular borders, necrotic centers) rather than artifacts.

Computational Efficiency:

• The refined model required fewer parameters and layers while maintaining high accuracy.
• Inference time increased only marginally (from 10ms to 11ms per image) due to the overhead of the XAI-guided training process, but the final model is structurally simpler.

5. Significance and Impact

• Trustworthy AI in Healthcare: By ensuring the model relies on tumor-relevant evidence, the approach directly addresses the "black box" problem, potentially increasing clinician trust and adoption of AI in radiology.
• Alignment with SDG 3: The work supports "Good Health and Well-being" by enabling more accurate, earlier, and non-invasive detection of brain tumors.
• Efficiency: Demonstrates that highly parameterized, complex models are not always necessary for high-performance medical imaging if the architecture is guided by explainability signals.
• Generalization: The strong performance on the unseen Dataset-2 suggests the model is robust to domain shifts (different scanners or acquisition protocols), a critical requirement for real-world clinical deployment.

In conclusion, this paper successfully bridges the gap between interpretability and architecture, proving that using explainability to prune and refine models leads to systems that are not only more accurate and efficient but also more clinically reliable.