Dr. Liming Zhang
University of Macau,  China

BIO: Liming Zhang received the B.S. degree in Computer Software at Nankai University, China and M.S. degree in Signal Processing at Nanjing University of Science and Technology, China. She received her PhD degree in image processing at University of New England, Australia. She is currently an assistant professor in Faculty of Science and Technology, University of Macau. Her research interests include Computer vision, Image processing, Artificial intelligence, Machine learning, and Deep learning. She has published over 100 papers, including IEEE Transactions on Image Processing, IEEE Transactions on Signal Processing, CVPR, ect. The main contribution lies on new image and signal processing methodology – adaptive Fourier decomposition (AFD)-based image processing methods and new deep network development. The image and video compression results based on stochastic AFD (SAFD) exceed the current international image and video compression standards JPEG, JPEG2000, MPEG, and also exceed the compression results of the popular deep networks. SAFD-based deep networks also perform well in alleviating the scarcity of training data.

Speech Title: A New Type of Self-supervised Learning Method Based on Mathematical Principle for Medical Image Segmentation

Abstract: In recent years, data-driven deep learning networks have achieved great success in medical image segmentation. However, these deep learning solutions often require large amounts of labeled data samples. This is especially difficult in the field of medical image processing, where well-annotated medical data is expensive and time-consuming. In literature, mathematical principles have demonstrated very good performance in unsupervised feature extraction. This paper introduces the recently developed mathematical theory – stochastic adaptive Fourier decomposition (SAFD) into the design of deep network to implement non-data-driven deep learning for medical image segmentation. The principle is to first use SAFD to self-supervised learn the convolution kernels of each layer, and then perform convolutions on the training images to extract deep features. Finally, a classifier is used together with one labeled training data for segmentation. Two non-pre-trained single-labeled data based training designs are proposed, including training with one labeled or one labeled and four unlabeled images. Extensive experiment results on six datasets from different modalities demonstrate the superiority of our proposed methods over other compared stateof-the-art fully supervised, semi-supervised and unsupervised methods, including recent popular Segment Anything Model (SAM) series. In addition, further generalization experiments in the same domain verify the robustness of the proposed method.