Shubham Gogri

ML Engineer

Overview

My MSc dissertation project focused on reducing radiation exposure in CT scans through deep learning-based image reconstruction. Medical diagnosis and radiotherapy treatment planning rely on the quantitative accuracy and quality of CT images, but reducing radiation dose and scanning time is critical for patient safety. Sparse-view CT—using fewer projection views—offers a solution, but introduces image quality degradation. This research applied deep learning to enhance sparse-view CT image quality.

The Problem

Computed Tomography (CT) imaging requires balancing two competing requirements: high image quality for accurate diagnosis, and minimizing radiation exposure to patients. Sparse-view CT reduces radiation by using fewer X-ray projections, but the incomplete projection data results in severe artifacts and lower image quality. The challenge: can deep learning reconstruct clinically usable images from sparse inputs?

Approach

Phase 1: MIR-Net Baseline

Initial experiments employed MIR-Net, a dedicated deep neural network designed for image enhancement. The architecture comprised encoder and decoder networks with Charbonnier Loss for optimization. While promising, this approach proved computationally demanding.

Phase 2: Pix2Pix GAN Architecture

A specialized GAN architecture rooted in the Pix2Pix framework was implemented:

Generator: U-Net-based architecture for image-to-image translation Discriminator: Convolutional Neural Network for adversarial training

Loss Functions (Key Innovation):

  • Charbonnier Loss: Captures minute details in reconstruction
  • Wasserstein Loss: Ensures training stability
  • Perceptual Loss: Calculated from VGG16 feature vectors (pretrained on ImageNet) to enhance synthesis of visually relevant images

Training Infrastructure

  • Containerized training pipelines using Docker
  • Deployed on University of Surrey’s HPC cluster (Condor system)
  • Comprehensive experiments with clinical CT data
  • Systematic exploration of loss function combinations

Results

The experiments demonstrated significant image quality improvements:

Metric Performance
SSIM Values approaching 1.0 (near-perfect structural similarity)
PSNR Up to 76 dB (excellent signal-to-noise ratio)

Additional findings:

  • Learning curves demonstrated enhanced training stability
  • Qualitative comparisons confirmed improved visual interpretation
  • Pixel value intensity preserved during reconstruction
  • GAN-based approaches significantly outperformed traditional methods

Recognition

  • Best Dissertation Nomination — University of Surrey, Electronic Engineering Industrial Advisory Board (2023-24)
  • Publication — Abstract accepted at ICMLMI 2023 (International Conference on Machine Learning in Medical Imaging), London

What I Learned

This project shaped how I think about applied ML research:

Technical insights:

  • Loss function design is as important as architecture design
  • Perceptual loss (VGG16 features) dramatically improves visual quality
  • Combining multiple loss functions requires careful balancing

Research methodology:

  • Containerization enables reproducible experimentation at scale
  • Systematic comparison across configurations reveals what actually works
  • Quantitative metrics (PSNR, SSIM) must be paired with qualitative assessment

Tech Stack

TensorFlow Pix2Pix GAN U-Net MIR-Net VGG16 Docker HPC/Condor Python OpenCV NumPy

Resources