Breast cancer represents a significant global health concern, emphasizing the importance of early detection and continuous monitoring to enhance patient outcomes. Traditional screening methods like mammography and tomosynthesis serve as essential tools in assessing breast structure and detecting suspicious nodules. However, the emergence of microwave radiometry (MWR) introduces a promising addition to this diagnostic arsenal. MWR measures internal tissue temperature, influenced by cellular metabolic activity. Cancerous cells, which demand increased energy as they proliferate, exhibit higher metabolic rates, leading to elevated heat generation detectable by MWR. This temperature-based assessment complements structural evaluations by providing insights into tumour growth rates. By integrating both approaches, clinicians can enhance breast cancer detection accuracy and monitoring, ultimately benefiting patient care. We have collected MWR data from >20,000 patients.
Nevertheless, the practical challenges of evaluating multiple imaging modalities within clinical workflows necessitate innovative solutions. Neural network algorithms offer a compelling avenue by seamlessly integrating diverse data modalities, thereby enhancing diagnostic accuracy while reducing clinicians’ workload. This project focuses on developing multimodal neural networks that integrate information from mammography, tomosynthesis, and MWR measurements. By leveraging the strengths of each modality, these networks aim to improve the efficiency and effectiveness of breast cancer detection. Moreover, we employ contrastive and self-contrastive learning techniques to enhance the model’s generalizability. While learning differences across patients is crucial, rich information lies within individual scans and measurements. Specifically, we exploit symmetrical properties between the breasts, enabling the network to identify subtle differences and similarities within corresponding regions. This multimodal contrastive approach not only promises improved diagnostic accuracy but also facilitates personalized treatment strategies. Furthermore, the techniques developed in this project have broader applicability to other anatomical regions such as the brain, showcasing their potential impact beyond breast cancer diagnosis.
The prospective student will join an interdisciplinary team focused on algorithmic, clinical, and MWR hardware advancements. They will gain hands-on experience with deep neural network algorithms for medical imaging and tackle real-world clinical challenges head-on.
We will consider including muti-omics and imaging data when they become available from collaborators.
The project is supported by industrial partner MMWR LTD, UK
