This project builds a deep learning-based chest X-ray (CXR) classification model to explore how AI can assist preliminary pneumonia screening.

πŸ“‹ Abstract

This project builds a deep learning-based chest X-ray (CXR) classification model to explore how AI can assist preliminary pneumonia screening.

Addressing the challenge that COVID-19 and pneumonia features are difficult to distinguish visually on X-rays, this project uses MATLAB as the development environment and implements ResNet50 for binary classification tasks (Normal vs. Pneumonia/COVID-19). The model showed relatively high recall on the test set, making it a useful research example for preliminary screening workflows.

πŸ›‘ The Problem

During pandemic peaks or in resource-limited regions, radiologists face enormous reading pressure. Traditional manual interpretation has the following pain points:

  1. Time-Consuming: Physicians need to focus for extended periods examining subtle pulmonary infiltration features.
  2. Ambiguous Features: Early pneumonia or COVID-19 ground-glass opacities closely resemble normal lung textures, easily causing misdiagnosis.
  3. Fatigue Errors: Visual fatigue from prolonged work may increase missed diagnosis rates.

πŸ› οΈ Technical Deep Dive

1. Data Preprocessing

The dataset contains two classes of chest X-ray images:

  • Normal: Clear lung fields with no significant abnormal shadows.
  • Pneumonia (Pneumonia/COVID-19): Lungs showing patchy shadows or consolidation.

To improve model generalization, I standardized the original images and resized them to ResNet50’s required input dimensions ($224 \times 224$ pixels).

Normal Chest X-Ray Figure 1: Normal chest X-ray image with clear lung fields.

Pneumonia Chest X-Ray Figure 2: Chest X-ray image showing pneumonia with visible fuzzy shadows in the lungs.

2. Core Architecture: ResNet50 and Residual Learning

Traditional Convolutional Neural Networks (CNN) often encounter vanishing gradient or gradient explosion problems when layers deepen, causing training performance to actually decrease.

To train a network as deep as 50 layers, I adopted the ResNet (Residual Network) architecture. Its core innovation is introducing “Shortcut Connections”.

Assuming the input to a neural network layer is $x$ and the desired feature mapping is $H(x)$. ResNet doesn’t directly learn $H(x)$, but learns the residual function $F(x) = H(x) - x$. The final output is:

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

  • $F(x, {W_i})$: Residual Mapping, the change the network layers need to learn.
  • $x$: Identity Mapping, directly passing input information to output.

This $y = F(x) + x$ structure ensures that even in extremely deep networks, gradients can flow back directly through this “highway” of $x$, allowing the model to effectively capture extremely subtle pathological features in X-ray images.

3. Training Strategy

Training was performed using MATLAB’s Deep Learning Toolbox.

  • Optimizer: SGDM (Stochastic Gradient Descent with Momentum)
  • Loss Function: Cross Entropy Loss (Binary Cross Entropy)

πŸ“Š Results & Analysis

Based on the test set Confusion Matrix, we quantitatively evaluated the model’s performance:

Confusion Matrix Results Figure 3: Test set Confusion Matrix.

1. Confusion Matrix Data Interpretation

Total test set samples: $N = 591$ images

  • True Positive (TP, Successfully predicted pneumonia): 308 images
  • True Negative (TN, Successfully predicted normal): 150 images
  • False Positive (FP, Incorrectly classified as pneumonia): 84 images
  • False Negative (FN, Missed pneumonia): 49 images

2. Key Metrics Calculation

  • Accuracy: $$Accuracy = \frac{TP + TN}{Total} = \frac{308 + 150}{591} \approx 77.5%$$ The model’s overall correct judgment rate is approximately 77.5%.

  • Recall / Sensitivity: $$Recall = \frac{TP}{TP + FN} = \frac{308}{308 + 49} \approx 86.3%$$ Recall is an important metric in medical-image screening. Among all actual pneumonia samples, the model identified 86.3% of cases, suggesting some value for reducing missed positives in a screening-oriented workflow.

  • Precision: $$Precision = \frac{TP}{TP + FP} = \frac{308}{308 + 84} \approx 78.6%$$ Among cases the model classified as pneumonia, 78.6% were confirmed. Although some normal individuals were misclassified as pneumonia (84 false positive cases), in medical screening contexts, “better to over-diagnose (False Positive) than to miss (False Negative)” is an acceptable trade-off.

Conclusion

This project implemented automated chest X-ray classification using the ResNet50 architecture. On the test set, the model reached 86.3% recall, which makes it a research prototype for flagging potentially high-risk images. Practical medical use would still require larger datasets, clinical validation, and integration with professional radiology workflows.