Dental OPG Image Synthesis using GAN

This repository contains a complete pipeline for synthesizing dental OPG images using a Generative Adversarial Network (GAN). The project includes two major components:

1. Data Preprocessing:
   Loads raw dental OPG images from a specified directory, resizes them to 256×256 pixels, normalizes them to the range [-1, 1], computes basic statistics, and saves the preprocessed images.

2. GAN Training:
   Trains a DCGAN-inspired architecture (with a Generator and a Discriminator) on the preprocessed images. The network is designed to generate 256×256 synthetic images. The training loop includes label smoothing, loss logging, and periodic saving of generated samples and evaluation plots.

Table of Contents

• Overview
• Features

Overview

The aim of this project is to generate realistic synthetic dental OPG images using GANs. The pipeline consists of:

• A preprocessing script (Cell 1) that:

  • Reads raw images from your Google Drive.
  • Resizes them to a fixed resolution of 256×256.
  • Normalizes the images (so pixel values lie in [-1, 1]).
  • Computes and prints useful metrics (number of images, average dimensions, per-channel mean, and standard deviation).
  • Saves the preprocessed images for further use.

• A GAN training script (Cell 2) that:

  • Loads the preprocessed images using torchvision.datasets.ImageFolder.
  • Defines a Generator and a Discriminator (both designed for 256×256 images).
  • Uses label smoothing and noise injection for robust adversarial training.
  • Logs and saves training losses and evaluation metrics.
  • Saves generated sample images at each epoch and produces plots of loss curves, evaluation metrics (dummy FID and Inception Score), and side-by-side comparisons of real vs. generated images.

Features

• Data Preprocessing:

  • Resizes images to 256×256 pixels.
  • Normalizes images to [-1, 1].
  • Computes statistics (e.g. average width, average height, per-channel mean and std).
  • Saves processed images for downstream use.

• GAN Architecture:

  • Generator256: Upsamples a latent vector (nz = 50) through 7 stages to generate 256×256 images.
  • Discriminator256: Processes 256×256 images through several convolutional layers to output a single probability (using a Sigmoid activation).

• Training Loop:

  • Implements adversarial training with label smoothing.
  • Logs generator and discriminator losses.
  • Saves generated image samples at each epoch.
  • Plots loss curves and evaluation metrics (FID and Inception Score placeholders).