A Research On Lesion Segmentation Based On Transfer Learning

The deep learning-based lesion segmentation algorithm is widely used in the field of medical image analysis,enabling doctors to make more accurate,faster decisions on disease diagnosis,treatment,and surgery.However,deep models heavily rely on high-quality data labels,and their performance drops drastically when tested on different datasets.Unsupervised domain adaptation algorithms have been proposed as solutions to these problems,mainly divided into two methods: adversarial learning-based and self-trainingbased.However,these two approaches have certain limitations;the former is difficult to train,and the latter often introduces noise due to domain shift issues.To address these issues,this thesis proposes a Cross-Pseudo Supervision Unsupervised Domain Adaptation(CPS-UDA)method based on a cross-pseudo supervision mechanism in Chapter 3.This method combines the ideas of consistency regularization and self-training,enforcing consistency regularization constraints on the pseudo-labels generated by two equivalent student models.This encourages the two models to interact with more useful information,thereby improving the model’s generalization performance on the target domain.Additionally,this method has two variants,namely CPS-UDADice and CPS-UDANR-Dice,which use Dice loss and NR-Dice loss,respectively,to implement the cross-pseudo supervision loss.The Dice loss can handle the extreme imbalance between the background and lesion regions in lesion segmentation tasks,while the NRDice loss reduces sensitivity to noise.In Chapter 3,this thesis compares four existing algorithms and validates the effectiveness of CPS-UDA from quantitative and qualitative perspectives.The algorithm proposed in Chapter 3 of this thesis overlooks the similarity information between the source and target domains in medical images.To fully leverage the potential similarity information between domain data and further improve the algorithm proposed in Chapter 3,this thesis introduces the Similarity-Based Cross-Pseudo Supervision Lesion Segmentation(S-CPS-UDA)algorithm based on the similarity of the HOG domain in Chapter 4.This algorithm uses the HOG feature descriptor to extract features from the source and target domains and calculates the similarity factor between each source domain sample and the target domain.Finally,this similarity factor is used to adaptively weight the supervised part of the source domain samples.The effectiveness of the similarity-weighted mechanism proposed in Chapter 4 of this thesis depends on the accuracy of feature extraction.Deep features often have better representational power and generalization performance than shallow features.Therefore,to further improve the algorithm proposed in Chapter 4,this thesis presents the Deep SimilarityBased Cross-Pseudo Supervised UDA Lesion Segmentation(Deep-S-CPS-UDA)algorithm based on the similarity of the deep domain in Chapter 5.This algorithm uses a pre-trained U-Net encoder on the source domain to extract features from the source and target domains,which are then used for calculating the similarity factor between each source domain sample and the target domain.Finally,through unsupervised domain adaptation experiments on three different lesions,this thesis demonstrates that the comprehensive performance of the S-CPS-UDA and Deep-CPS-UDA methods based on the similarity-weighted mechanism is superior to the CPS-UDA method proposed in Chapter 3.