How yash-vks-chauhan/ct-denoising-u-net works
Project Type & Purpose
• This project is a deep learning-based image processing system focused on denoising lung CT scans and X-ray images.
• It includes a U-Net model implementation built with TensorFlow/Keras, accompanied by a preprocessing pipeline, model training (via train.py), evaluation metrics (PSNR, SSIM), and even a notebook-based exploration.
Analyzing the File Structure
• Key files and directories include:
• README.md: Provides an overview of the project, features, and instructions.
• train.py: The training script that initiates model training.
• denoising_model.h5 & denoising_metrics.csv: Represent the trained model and evaluation metrics respectively.
• notebooks folder: Contains Jupyter notebooks for testing and training analysis.
• are folder: Contains sample images such as lung_images.png.
• GitHub workflow files under .github: Show CI/CD integrations (e.g., pylint for code quality).
Main Components
• Data Input and Dataset Handling:
• Inputs include CT scans from various datasets (e.g., COVID-19 Radiography, NIH Chest X-ray).
• Separate input folders for “Noisy” and “Clean” images, as mentioned in the README.
• Preprocessing Pipeline:
• Steps like grayscale conversion, resizing, normalization, CLAHE, augmentation, and possibly super-resolution modifications.
• Conversion of raw images into a standardized format for training.
• U-Net Denoising Model:
• Core deep learning model architecture which consists of:
• Encoder (series of convolutional layers)
• Bottleneck (capturing noise patterns)
• Decoder (reconstruction of denoised images)
• Training Process:
• Training inputs (paired noisy and clean images) feed into the model.
• Visualization components, such as the training history and evaluation curves (loss, MAE) that are recorded.
• Model output (denoising_model.h5) and metrics (denoising_metrics.csv).
• Inference and Evaluation:
• Steps where inference.py may be used for processing new images.
• Evaluation metrics computation for PSNR and SSIM.
• Validation pipeline confirmation using augmented vs. original image consistency.
Relationships & Interactions
• From raw image data entering the preprocessor.
• From the preprocessed dataset feeding into the U-Net model training.
• From the training pipeline towards evaluation metrics and model output.
• Interactions between batch data generators (as mentioned in the README) and GPU computations for accelerated training.
Architectural Patterns & Technology Stack
• Clearly delineate the separation between data handling (preprocessing, augmentation) and model architecture (encoder, bottleneck, decoder).
• Use of TensorFlow/Keras for deep learning and note that GPU acceleration is leveraged.
• Mention the use of custom data generators for efficient memory management.
• CI/CD integration shown by the .github/workflows configuration which ensures code quality.