Cross-level Convolutional Transformer and Adversarial Multi-task Learning for Medical Semantic Segmentation

点击翻译按钮获取中文摘要

This project plans to study cross-level convolutional Transformers and adversarial multi-task learning for medical semantic segmentation (MSS). The goal of MSS labels each pixel of an image with a corresponding class of what is being represented to provide the segmentation maps. Though deep learning-based methods have achieved state-of-the-art performance in MSS, they still struggle to achieve fine-grained segmentation maps in complex environments, which prohibits their implementation to real-world applications.To handle this problem. I plan to explore some promising algorithms for MSS. First, I will focus on improving skip connections of UNet by proposing a convolutional Transformer with cross-level interaction. Second, I will aim to break the shackle of model performance caused by the small number of annotations, through a shared-private architecture with adversarial multi-task learning to use as much additional data as possible.The potential impact of this project mainly includes two aspects: First, the methods studied in this project will improve the model performance of MSS in complex scenes and have strong potential to broaden the applicability of multi-modal/unlabeled/multi-task data. Second, this project will develop a generalized and instructive structure for MSS thanks to the shared-private mechanism and adversarial learning. It could be used for multi-tasks with heterogeneous inputs, with only a few modifications.Aims and ObjectivesThis project aims to address two key challenges to improve the performance and usability of MSS in complex surgical environments.Challenge 1: How to extract high-quality features and fuse them effectively?The latest MSS methods based on UNet fail to explore sufficient information from full scales due to the following two aspects:a) Not all connection pathways are effective due to the issue of semantic gaps in different layers. Those redundant and irrelevant connections increase the training difficulty of the network, even some can undermine the performance.b) The optimal combination of skip contributions is varied among different datasets, which depends on the scales and appearance of segmentation objects.To address the above problems, I consider replacing vanilla skip-connection pathways with Transformers to capture non-local features and perform effectively cross-level feature fusion.Challenge 2: How to utilize multi-modal data, unlabeled data, or even data from other tasks to improve the model performance?The scarcity of carefully-labelled datasets becomes an unavoidable limitation in DL-based MSS as both data and annotations are expensive to acquire. Previous methods pay less attention to utilizing different types of external data. Therefore, I consider using adversarial multi-task learning to build a uniform architecture for additional data, regardless of its type. This project will propose a novel cross-level convolutional Transformer for MSS to improve the skip-connection process of UNet.This project will propose a novel shared-private network with multiple encoders and decoders for MSS to utilize adversarial multi-task learning to handle additional data or tasks.As surgical image data are generally characterized by multiple modalities/scales and complex scenes, the research based on it could shed light on complex MSS.