Pneumonia Prediction in Chest X-Ray Images

Medical Context

Pneumonia is a lung infection that inflames the air sacs in one or both lungs, filling them with fluid or pus. This causes symptoms such as cough with phlegm, fever, chills, and difficulty breathing. Various microorganisms (bacteria, viruses, and fungi) can cause it.

In a chest X-ray, the key differences are:

Healthy Lung	Lung with Pneumonia
Lung transparency (appear dark due to air)	Opacities or consolidations (white/gray areas due to fluid or pus)
No abnormal white areas	Air bronchogram (dark lines within white areas)
Defined structures (heart, diaphragm, vessels)	Lobar or interstitial infiltrates

---

DenseNet-201

DenseNet-201 is a deep convolutional neural network with 201 layers, belonging to the Dense Convolutional Networks (DenseNets) family, introduced in 2016 by Gao Huang and collaborators.

In a traditional convolutional network with L layers there are L connections (each layer connects only to its successor). In a DenseNet with L layers there are L(L+1)/2 direct connections: each layer receives as input the feature maps of all previous layers and delivers its own to all subsequent layers. This enables more efficient gradient flow and reduces the vanishing gradient problem.

---

Dataset

The Pneumonia X-Ray Images dataset by Paulo Breviglieri (Kaggle) was used, a more balanced version of the original dataset by Paul Mooney.

Image distribution:

Split	Pneumonia	Normal	Total
Train	3,110	1,082	4,192
Validation	773	267	1,040
Test	390	234	624

A significant imbalance is observed: the “Pneumonia” class has approximately three times as many images as the “Normal” class in each split.

---

Preprocessing

• Resized to 224×224 pixels with 3 RGB channels
• Pixel normalization (rescaling 1/255)
• Data augmentation applied to the training generator:
  • Random rotation (±30°)
  • Random zoom (0.2)
  • Horizontal and vertical shift (0.2)
  • Brightness variation (range 0.7–1.3)
  • Random horizontal flip
  • Nearest-neighbor fill

---

Model Architecture

Transfer learning was applied using DenseNet-201 pre-trained with ImageNet weights as a feature extractor (frozen layers), with custom top layers added:

Input (224×224×3)
  └── DenseNet-201 (frozen, ImageNet weights)
        └── GlobalAveragePooling2D
              └── Dense(1536, relu) → Dropout(0.1)
                    └── Dense(512, relu) → Dropout(0.1)
                          └── Dense(1024, relu) → Dropout(0.1)
                                └── Dense(512, relu) → Dropout(0.1)
                                      └── Dense(2, softmax)

The layer configuration and dropout (0.1) were based on the results from the paper “Optimizing DenseNet for Potato Leaf Disease Classification” (arXiv 2402.03347), which showed that a dropout of 0.1 achieves the best balance between training and validation performance.

---

Training

Parameter	Value
Optimizer	Adam
Loss function	Categorical Crossentropy
Metric	Accuracy
Max epochs	20
Early Stopping	Patience of 6 epochs, monitoring val_loss
Batch size	32 (training), 8 (test)

Class weights were computed inversely proportional to frequency, but were not applied during training.

---

Results

Training curve

The model reached ~95–96% accuracy on both training and validation over the 20 epochs, with decreasing losses and negligible overfitting between both curves.

Test evaluation

Metric	Normal Class (0)	Pneumonia Class (1)
Precision	0.95	0.80
Recall	0.59	0.98
F1-score	0.73	0.88

Global test accuracy: 83.8% (loss: 0.42)

Confusion Matrix

	Predicted Normal	Predicted Pneumonia
Actual Normal	135	99
Actual Pneumonia	6	384

Interpretation

• The model correctly detects most pneumonia cases (98% recall).
• However, it has poor performance on the Normal class (only 59% recall), misclassifying 99 normal images as pneumonia.
• This reveals a bias toward the dominant class due to dataset imbalance and the non-application of the computed class weights.

---

Conclusions

1. High accuracy on pneumonia — the model correctly identifies most positive cases.
2. Low performance on normal class — high false positive rate (99 images) due to class imbalance.
3. Overfitting — although validation showed good performance, the test revealed a significant drop, possibly because train and validation came from the same ImageDataGenerator, which may have led to pattern memorization.

Proposed Improvements

1. Apply class_weights during training to balance the importance of both classes.
2. Fine-tuning by unfreezing top layers of DenseNet-201 with a reduced learning rate.
3. Targeted data augmentation for the minority class (Normal) to improve generalization.

---

Versions

The project was developed with the following libraries and versions:

Package	Version
Python	3.12.3
TensorFlow	2.18.0
Keras	3.7.0
scikit-learn	1.5.1
matplotlib	3.9.2
numpy	1.26.4
pandas	2.2.2
seaborn	0.13.2

---

References

• Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely Connected Convolutional Networks. arXiv:1608.06993.
• Breviglieri, P. Pneumonia X-Ray Images. Kaggle.
• Mooney, P. Chest X-Ray Images (Pneumonia). Kaggle.
• Reference paper for adapted architecture: arXiv 2402.03347.