Transfer Learning from Medical Ultrasound to Marine Sonar Image Data

Project Description:

This project explores using medical ultrasound (US) image data to train neural networks for marine sonar applications. Through transfer learning approaches - including zero-shot, one-shot, and few-shot learning - we aim to determine if knowledge from medical ultrasound images can enhance 2D sonar data analysis.

Motivation:

Due to the scarcity of sonar image data, training neural networks from scratch remains challenging. This scarcity limits the development of robust models for sonar-specific tasks. Medical ultrasound images, however, are more widely available and operate on similar acoustic imaging principles. By leveraging medical ultrasound data for training, we can expand our dataset substantially, leading to more effective training and better neural network performance on sonar data. This approach could reduce our dependence on large sonar-specific datasets while promoting innovation across domains.

Key areas of focus include:

1. Training Neural Network: Implementing and applying a training pipeline to train a neural network for an application such as object detection or segmentation

2. Transfer Learning Feasibility: Testing if models trained on medical ultrasound data can effectively process sonar images

3. Impact Analysis: Comparing the performance of medical US-trained models to traditional sonar-specific approaches

4. Applications: Exploring practical uses in sonar image object detection and segmentation for underwater exploration and monitoring

Required skills:

• Proficiency in Python programming

• Experience with deep learning frameworks such as Keras, TensorFlow, or PyTorch

• Solid understanding of deep learning concepts, including convolutional neural networks (CNNs)

• English or German speaking and writing skills

This research bridges medical and marine imaging domains to reduce dependency on large sonar datasets while advancing neural network adaptability across imaging types. The findings could offer new solutions for both medical imaging and marine applications.