Enhancing Pneumonia Diagnosis with Deep Neural Networks

Developed an intelligent medical image classification system to assist in pneumonia detection using chest X-ray images, focusing on building a reliable and generalizable deep learning solution for real-world healthcare applications. The project was developed through multiple iterative cycles, addressing practical challenges such as class imbalance, overfitting, and inconsistent model predictions.

Key Contributions & Technologies:

● Evaluated deep learning architectures including DenseNet121, ResNet50, and EfficientNet-B3, and refined the approach based on performance and generalization challenges.
● Designed and implemented a custom CNN model, applying data augmentation, dropout regularization, L2 weight decay, and class weighting to improve model performance.
● Applied early stopping and optimized prediction thresholds using F1-score analysis to ensure balanced and reliable classification.
● Achieved 88.2% accuracy, 90.4% precision, and 90.3% recall, outperforming initial benchmark models.
● Integrated the model into a Streamlit-based web application for real-time image upload and prediction with confidence scores.

Research & Development Approach:

● Conducted iterative experimentation to address model instability and improve generalization.
● Identified limitations in transfer learning approaches and refined the architecture accordingly.
● Focused on developing a practical and deployable AI solution, rather than only optimizing theoretical performance.

Tools & Technologies:

● Python, TensorFlow, Keras
● NumPy, Pandas, Matplotlib, Scikit-learn
● Deep Learning, CNN, Medical Image Processing
● Streamlit (Deployment)

Outcome & Future Plans:

Successfully developed a reliable pneumonia detection system with practical deployment capabilities. Future work includes improving model interpretability, integrating attention mechanisms, and extending the system for multi-disease classification.