⚡ electrical engineering

CBAM-Enhanced DenseNet121 for Multi-Class Chest X-Ray Classification with Grad-CAM Explainability

This paper introduces CBAM-DenseNet121, a transfer-learning framework enhanced with the Convolutional Block Attention Module that achieves high accuracy and interpretability in distinguishing between normal, bacterial, and viral pneumonia in chest X-rays, addressing the critical need for multi-class diagnosis in resource-constrained settings while also revealing that EfficientNetB3 underperforms a custom CNN baseline in binary tasks.

Original authors: Utsho Kumar Dey

Published 2026-04-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Utsho Kumar Dey

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor in a busy hospital in Bangladesh. You have a stack of chest X-rays on your desk, and you need to figure out what's wrong with each patient. But here's the catch: you are tired, there are too many patients, and sometimes the X-rays look very similar even though the diseases are different.

This paper is about building a smart computer assistant to help you sort through these X-rays, but with a special twist: it doesn't just say "Sick" or "Not Sick." It tells you exactly what kind of sickness it is.

Here is the story of how they built this assistant, explained in simple terms:

1. The Problem: The "Sick" vs. "Not Sick" Trap

Most computer programs that look at X-rays are like a simple security guard at a door. They only ask two questions: "Is this person healthy?" or "Is this person sick?"

The Flaw: If the guard says "Sick," you don't know why. Is it a Bacterial infection (which needs antibiotics) or a Viral infection (which needs rest and fluids, but no antibiotics)?
The Risk: If you give antibiotics for a virus, it's useless and can make bacteria stronger in the future (a big problem called "superbugs").
The Goal: The authors wanted a computer that acts like a specialist, distinguishing between Normal, Bacterial Pneumonia, and Viral Pneumonia.

2. The Solution: The "Super-Brain" with Glasses

The team built a new model called CBAM-DenseNet121. Let's break down what that means using an analogy:

The Base (DenseNet121): Imagine a very smart student who has already read a million books about general pictures (like cats, cars, and trees). This student is our "base" AI. They are already pretty good at seeing shapes.
The Glasses (CBAM): The student is smart, but they need better glasses to focus on the tiny details in an X-ray. The authors added a special pair of "attention glasses" called CBAM.
- Channel Attention: This helps the AI decide which features matter. It's like telling the student, "Ignore the background noise; focus only on the lung textures."
- Spatial Attention: This helps the AI decide where to look. It's like pointing a finger and saying, "Look right here, in the bottom corner of the lung, that's where the problem is."
The Result: By putting these glasses on the smart student, the AI becomes much better at spotting the subtle differences between bacterial and viral infections.

3. The "Negative" Discovery: Bigger Isn't Always Better

Before building their final model, the team tested other famous AI models to see which one was best.

The Surprise: They found that a very popular, complex model called EfficientNet actually did worse than a simple, custom-made model they built from scratch.
The Lesson: It's like bringing a Formula 1 race car to a muddy farm field. The race car is amazing on a track, but it gets stuck in the mud. Similarly, some fancy AI models are trained on colorful photos of nature, but they get confused by the gray, grainy texture of medical X-rays. Sometimes, a simpler, specialized tool works better.

4. Trusting the Machine: The "Flashlight" Test

Doctors are often scared of AI because it feels like a "black box"—you put an image in, and a number comes out, but you don't know why.

The Fix: The team used a tool called Grad-CAM. Think of this as a flashlight.
How it works: When the AI says, "This is Viral Pneumonia," the flashlight shines a red glow on the exact part of the X-ray that made the AI think that.
The Proof: When they looked at the glowing spots, they matched perfectly with what human doctors look for.
- For Bacterial cases, the light glowed on solid, white patches (like a heavy blanket).
- For Viral cases, the light glowed on fuzzy, cloudy areas (like fog).
- For Normal cases, the light was spread out evenly, showing no specific trouble spots.
  This proves the AI isn't just guessing; it's actually "seeing" the disease.

5. The Results: Reliable and Repeatable

To make sure their results weren't just a lucky fluke, they ran the test three times with different random settings (like rolling dice three times to make sure the game is fair).

The Score: Their new "Super-Brain" got about 84% accuracy in telling the three categories apart.
The Stability: The scores were almost identical every time they ran the test, proving the system is stable and reliable.

Why Does This Matter?

In places like rural Bangladesh, there aren't enough expert doctors to look at every X-ray. This system acts as a first-line triage nurse.

It looks at the X-ray.
It highlights the problem area with a "flashlight."
It tells the doctor: "This looks like a bacterial infection; you might need antibiotics," or "This looks viral; let's hold off on drugs."

It doesn't replace the doctor; it gives the doctor a second opinion and a visual map, making healthcare faster, safer, and more accurate for everyone.

1. Problem Statement

Pneumonia is a leading cause of childhood mortality, particularly in low-resource settings like Bangladesh where access to radiologists is limited. While deep learning has advanced pneumonia detection, most existing models treat the task as a binary classification problem (Normal vs. Pneumonia). This approach overlooks a critical clinical distinction:

Bacterial Pneumonia: Requires antibiotic treatment.
Viral Pneumonia: Requires supportive care.
Consequence: Conflating these two types can lead to unnecessary antibiotic prescriptions, exacerbating antimicrobial resistance.

Furthermore, many medical AI models lack interpretability, making it difficult for clinicians to trust automated diagnoses. The paper addresses the need for a three-class classification system (Normal, Bacterial, Viral) that is both accurate and explainable.

2. Methodology

A. Dataset and Preprocessing

Source: The Kermany (Chest X-Ray Images) dataset from Kaggle, containing 5,863 images.
Reorganization: The authors discarded the original, statistically insufficient validation split (16 images) and applied stratified random sampling (80/20) to the training folder.
- Training/Validation: 5,216 images (split into 4,173 train / 1,043 val).
- Test Set: 624 images (held out completely).
- Classes: Normal (1,341), Bacterial Pneumonia (2,530), Viral Pneumonia (1,345).
Class Imbalance: Addressed using balanced class weights (inverse-frequency scaling) applied to the categorical cross-entropy loss.
Augmentation: Applied only to training data (rotation, flipping, shifting, zoom, brightness variation).

B. Proposed Architecture: CBAM-DenseNet121

The core model integrates the Convolutional Block Attention Module (CBAM) into a DenseNet121 backbone (pre-trained on ImageNet).

Backbone: DenseNet121 provides dense skip connections to mitigate vanishing gradients and promote feature reuse.
CBAM Integration: Inserted after the final convolutional block (7×7×1024) and before global average pooling. It operates in two sequential steps:
1. Channel Attention: Identifies which feature maps are important by using global average/max pooling and a shared MLP.
2. Spatial Attention: Identifies where to focus within those maps by concatenating channel-wise pooled maps and applying a 7×7 convolution.
Classifier Head: GlobalAveragePooling → BatchNorm → Dense(512, ReLU) → Dropout(0.5) → Dense(256, ReLU) → Dropout(0.3) → Softmax(3).

C. Training Strategy

A two-phase training strategy was employed to prevent catastrophic forgetting:

Phase 1 (Freezing): The DenseNet121 base is frozen; only CBAM and the classifier head are trained for 15 epochs (Learning Rate: $10^{-3}$ ).
Phase 2 (Fine-tuning): The last 30 layers of DenseNet121 are unfrozen and trained for 20 epochs (Learning Rate: $10^{-5}$ ).

Optimization: Adam optimizer with Early Stopping (patience=5) and ReduceLROnPlateau.

D. Baselines and Validation

Binary Baselines: To establish reference points, four models were tested on the binary task (Normal vs. Pneumonia): Custom CNN, ResNet50, EfficientNetB3, and standard DenseNet121.
Statistical Robustness: All experiments were repeated three times with independent random seeds (42, 7, 123). Results are reported as Mean ± Standard Deviation to ensure reproducibility.
Explainability: Grad-CAM was applied to the final convolutional layer to visualize the regions influencing the model's predictions.

3. Key Contributions

Three-Class Framework: Moves beyond binary detection to distinguish Normal, Bacterial, and Viral pneumonia, offering actionable diagnostic output.
CBAM Integration: Successfully integrates dual-axis attention (channel and spatial) into DenseNet121 to enhance feature discrimination in medical images.
Negative Finding on Architecture Selection: Demonstrated that EfficientNetB3 (73.88%) underperformed a custom CNN baseline (78.53%) on this grayscale medical task, challenging the assumption that larger, newer architectures always yield better results without domain-specific pretraining.
Statistical Rigor: Validated results across three independent runs with low standard deviation, confirming training stability.
Interpretability: Provided Grad-CAM visualizations showing the model focuses on anatomically plausible lung regions.

4. Experimental Results

A. Binary Task Baseline (Reference Only)

DenseNet121 achieved the highest accuracy (91.03%) and AUC (0.9706).
EfficientNetB3 performed poorly (73.88%), underperforming even the simple custom CNN.
ResNet50 showed the highest recall for pneumonia (0.95), suitable for high-sensitivity screening.

B. Three-Class Performance (CBAM-DenseNet121)

Overall Accuracy: 84.29% ± 1.14%.
Macro-Average AUC: 0.9454 ± 0.0013.
Per-Class AUC:
- Normal: 0.9610 ± 0.0014 (Excellent)
- Bacterial Pneumonia: 0.9565 ± 0.0010 (Excellent)
- Viral Pneumonia: 0.9187 ± 0.0037 (Good)
Confusion Matrix Analysis:
- Bacterial pneumonia was classified with high accuracy (93.4%), likely due to distinct lobar consolidation patterns.
- The primary error mode was confusing Viral Pneumonia with Normal cases (32 false negatives), attributed to the subtle nature of early viral infiltrates.

C. Explainability (Grad-CAM)

Visualizations confirmed that the model attends to:

Bacterial: Lower-lobe consolidation zones.
Normal: Mediastinal regions without focal pathology.
Viral: Diffuse peripheral regions (interstitial patterns).

5. Significance and Conclusion

This paper presents a robust, reproducible, and interpretable framework for multi-class pneumonia detection.

Clinical Impact: By distinguishing between bacterial and viral causes, the model supports better treatment decisions (antibiotics vs. supportive care), directly addressing the issue of antimicrobial resistance.
Deployment: The model is lightweight enough for standard JPEG inputs, making it suitable for resource-constrained environments like rural Bangladesh.
Research Insight: The study provides a crucial cautionary note for medical AI researchers: larger architectures (like EfficientNet) do not automatically outperform simpler baselines on low-contrast grayscale medical images without specific domain adaptation.
Future Work: The authors plan to validate on larger, multi-institutional datasets (CheXpert, NIH) and perform prospective clinical trials.

The low standard deviation across all metrics confirms that the proposed CBAM-DenseNet121 is a stable and reliable tool for potential clinical deployment under radiologist supervision.