Redesigning Transformer Neural Networks for Data-Efficient Breast Cancer Diagnosis

Transformer neural network models have achieved state of the art results in many areas of machine learning, including natural language processing (e.g. BERT, GPT-3) and computer vision (e.g. ViT, SAM). However, to achieve these results currently requires extremely large training datasets which are not available in every domain. One such domain is medical imaging, in which data is often limited due to privacy concerns and costs. This project aims to overcome this limitation by integrating domain-specific knowledge into a transformer pre-trained on large general datasets. In contrast to previous work, which generally treats the transformer as a black-box, our approach will integrate domain-specific knowledge by modifying the transformer architecture.

Faculty Supervisor:

Hamid Usefi

Student:

Partner:

Artinus Consulting Inc.

Discipline:

Computer science

Sector:

Professional, scientific and technical services

University:

Memorial University of Newfoundland

Program: