A Technical Study On Medical Image Segmentation Based On Deep Learning Method

Medical image segmentation is an essential component in the computer-aided diagnostic process and a key technology in automated medical image analysis.It provides crucial information on physical structures and tissues in images,assisting physicians in understanding disease conditions and thus has significant practical value.In recent years,it has become a research hotspot in the field of pattern recognition and computer vision.Compared to traditional segmentation methods based on thresholding,region growing,edge detection,and machine learning relying on feature engineering,deep learning-based medical image segmentation methods have received widespread attention and become mainstream with their superior performance and end-to-end design.This dissertation addresses the common challenges of deep learning-based medical image segmentation methods and proposes several new semantic segmentation methods and models as improved solutions.The main research results are summarized as follows:(1)Representation learning is the basis of semantic segmentation,and this paper proposes a pre-training method based on medical images to address the problem of visual semantic inconsistency between natural images and medical images in transfer learning.Firstly,a method to quantify image modal differences is proposed to explain the reasons for visual semantic inconsistency caused by modal differences,and propose a pre-training strategy based on mixed-modality medical images to solve the problem of insufficient visual information in medical image pre-training; Meanwhile,a weakly supervised learning representation transfer strategy and cross-scale semantic aggregation model,Hy Net,are proposed to address the problem of huge labour and time costs for sample annotation at the pixel level.(2)To address the problem of fine-grained and boundary semantic missing under the condition of limited computing resources.In this dissertation,a Fine-grained Boundaryenhancing Semantic Segmentation(FBSS)method based on metric learning is first proposed.The FBSS is based on a boundary-anchored adaptive triplet sampling strategy for explicit boundary semantic learning,which enables the preservation of fine-grained semantics on fuzzy boundaries.An Explicit Boundary-guided Knowledge Distillation(EBKD)method is also proposed,which uses boundary-guided deep supervision and boundary feature embedding alignment learning to explicitly migrate the boundary semantic information and effectively retain the boundary semantic information learnt from pre-training of large models.The customised use of metric learning and knowledge distillation methods avoids the introduction of additional computational effort while improving semantic richness.(3)To address the problems of missing pathological information and insufficient introduction of cross-modal semantics in small sample scenarios,existing mainstream methods can only implicitly learn pathological information based on image input,resulting in insufficient acquisition of pathological semantics and a significant decrease in the ability to localize and discriminate lesions.This dissertation proposes a Cross-modal Attentionbased Semantic Segmentation(CASS)method for medical images based on a cross-modal attention mechanism.CASS introduces text as model input data in addition to images,explicitly introduces pathology-related information such as lesion location and type,and adaptively refines the information within each modality and fuses the cross-modal semantics to solve the problem of insufficient pathology information.Based on the above research results,this dissertation proposes several solutions to the common challenges in semantic segmentation of medical images with deep learning.Extensive experiments on several publicly available semantic segmentation tasks show that the methods in this dissertation can effectively reduce the number of samples required for pre-training and human annotation costs,and avoid the semantic inconsistency problem.While not introducing additional inference computation,the methods effectively preserve the fine-grained boundary semantics and improves the performance of segmentation on fuzzy boundaries of medical images.And by introducing cross-modal inputs,the methods effectively compensate for pathological information and suppresses the appearance of false-positive regions in small sample scenarios,achieving the optimal comprehensive performance in the same period.