Title:	Breast Tomothsynthesis Generation and Reconstrcution
DNr:	Berzelius-2023-99
Project Type:	LiU Berzelius
Principal Investigator:	Zhikai Yang <zhikai@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2023-04-17 – 2023-11-01
Classification:	20603
Homepage:	https://bosomshield.eu/
Keywords:

Abstract

Breast cancer have been a leading cause of female mortality in the world. The early screening technique have been an prevalent technique to detect breast cancer. Especially the digital breast tomosynthesis have been an effective and efficient way to detection breast tumor. But due to the large image size and redundant information, radiologist need plenty of time to check the image. The computer aided diagnosis method could automatically check the image which could alleviate radiologist workload. Especially the deep learning based method. In recent years, there are amount of research results have been proven the deep knowledge based method have a unique advantage compared with the classic method in terms of prediction accuracy and generalization. 1.Self-supervised learning-based method, which pre-trained the neural network model with unlabeled data to learn a better-hidden representation. 2.Image synthesis-based approach, which generates more data for deep learning by using normal and tumor cases together. We plan to propose a deep learning based digital breast tomosynthesis diagnosis method which could alleviate radiologists' workload. But the performance of the deep learning model is heavily dependent on the dataset. In medical image analysis, the labeled data is hard to acquire, and the distribution of different diseases is different. In this work, we attempt to propose a new method that could employ normal case data to improve the model performance, which is essential to develop deep learning based computer-aided diagnosis system.