Deep Image Prior (DIP) for Fetal Ultrasound Enhancement in MATLAB

October 28, 2025 · View on GitHub

This repository contains the MATLAB implementation of the Deep Image Prior (DIP) method for solving inverse image problems such as denoising and super-resolution — specifically applied to fetal ultrasound images.

The project was developed as part of the MathWorks Excellence in Innovation program and also as the author's bachelor thesis at ETSEIB – UPC Barcelona.

Project Objectives

  • Adapt Python DIP code to MATLAB using native deep learning tools.
  • Implement DIP from scratch without pretraining or external datasets.
  • Evaluate DIP performance on ultrasound and standard test images.
  • Focus on medical imaging: improving the quality of fetal ultrasound images where clarity is crucial for diagnosis.
  • Explore its feasibility in low-resource environments (no GPU, no large datasets).

DIP uses an untrained convolutional neural network as an image prior. Unlike typical deep learning methods, it does not require any dataset or training phase — it fits a randomly initialized network to a single corrupted image.

Quick Start

Main Scripts

  • MAIN.mlx - Interactive MATLAB Live Script with the main workflow
  • runDenoising.m - Execute denoising tasks
  • runSuperResolution.m - Execute super-resolution tasks

Usage

  1. Open MAIN.mlx in MATLAB for an interactive experience
  2. Or run the specific task scripts directly: 
    runDenoising()      % For denoising tasks
runSuperResolution() % For super-resolution tasks

Tasks Implemented

  • Denoising - Apply DIP to remove Gaussian and blind noise from ultrasound and standard images
  • Super-resolution - Recover high-resolution detail from low-resolution fetal ultrasound images

Requirements

  • MATLAB R2023a or newer
  • Deep Learning Toolbox
  • Image Processing Toolbox
  • (Optional: use the Deep Network Designer app for visualization)

Repository Structure

Key Features

  • Modular Architecture: Clean separation of concerns with dedicated folders for different functionalities
  • Reusable Components: Common utilities shared across different tasks
  • Interactive Workflow: Live Script interface for easy experimentation
  • Medical Imaging Focus: Specialized for fetal ultrasound enhancement
  • No Training Required: Works with single images without datasets

References

  • Ulyanov, Vedaldi & Lempitsky, "Deep Image Prior", 2018.
    [Paper]
  • MathWorks DIP Challenge Repository
    GitHub Link
  • Montal Morta, M. (2025). Deep Image Prior applied to Fetal Ultrasounds.
    Bachelor's Thesis, ETSEIB – UPC Barcelona.

Author: Mariona Montal Morta

Bachelor in Industrial Technologies and Economic Analysis
UPC Barcelona – July 2025
Supervisor: Antoni Susín Sánchez

License

This project is open-source and shared under the MIT License.