A novel approach named hierarchical vascular morphology awareness and noisy label refine for hepatic vessel segmentation(HN-Seg)is proposed to achieve precise vessel segmentation while reducing dependency on high-quality labels.HN-Seg comprises of(1)hierarchical vascular morphology aware network which employs a multi-scale local morphology attention mechanism and a global morphology preservation loss function to ensure the integrity of overall vascular morphology,and(2)self-distillation noisy label refine module which leverages the uncertainty in model outputs to optimize noisy labels through uncertainty optimization and consistency regularization,thereby maximizing the knowledge extracted from images during training and refining noisy labels.Experimental results on the hepatic vessel dataset demonstrate that HN-Seg achieves superior segmentation performance,outperforming 6 methods(UNet,UNet++,UNETR,SwinUNetR,FRUNet,and MTCL).HN-Seg attains DSC and clDice scores of 0.727 and 0.773,showing improvements of 9.6%and 21.5%over the baseline method UNETR.

A self-supervised super-resolution reconstruction method based on scale adaptive and coordinate encoding is proposed to realize super-resolution reconstruction of anisotropic brain magnetic resonance images with different slice thicknesses even in the absence of paired isotropic brain magnetic resonance images.Firstly,an image encoding module that integrates super-resolution scale information is used to learn the specific features of images with different slice thicknesses.Subsequently,a coordinate encoding module is employed to facilitate the deep fusion of coordinate information and image features.Finally,an overall loss function comprising reconstruction loss and brain tissue edge perception loss is adopted to optimize the recovery of edge high-frequency information,while global residual learning is introduced to enhance network training.Experimental results on the HCP-1200 and OASIS-1 datasets demonstrate that the proposed method outperforms other self-supervised super-resolution reconstruction methods.

A self-supervised super-resolution reconstruction method based on scale adaptive and coordinate encoding is proposed to realize super-resolution reconstruction of anisotropic brain magnetic resonance images with different slice thicknesses even in the absence of paired isotropic brain magnetic resonance images.Firstly,an image encoding module that integrates super-resolution scale information is used to learn the specific features of images with different slice thicknesses.Subsequently,a coordinate encoding module is employed to facilitate the deep fusion of coordinate information and image features.Finally,an overall loss function comprising reconstruction loss and brain tissue edge perception loss is adopted to optimize the recovery of edge high-frequency information,while global residual learning is introduced to enhance network training.Experimental results on the HCP-1200 and OASIS-1 datasets demonstrate that the proposed method outperforms other self-supervised super-resolution reconstruction methods.

A novel approach named hierarchical vascular morphology awareness and noisy label refine for hepatic vessel segmentation(HN-Seg)is proposed to achieve precise vessel segmentation while reducing dependency on high-quality labels.HN-Seg comprises of(1)hierarchical vascular morphology aware network which employs a multi-scale local morphology attention mechanism and a global morphology preservation loss function to ensure the integrity of overall vascular morphology,and(2)self-distillation noisy label refine module which leverages the uncertainty in model outputs to optimize noisy labels through uncertainty optimization and consistency regularization,thereby maximizing the knowledge extracted from images during training and refining noisy labels.Experimental results on the hepatic vessel dataset demonstrate that HN-Seg achieves superior segmentation performance,outperforming 6 methods(UNet,UNet++,UNETR,SwinUNetR,FRUNet,and MTCL).HN-Seg attains DSC and clDice scores of 0.727 and 0.773,showing improvements of 9.6%and 21.5%over the baseline method UNETR.

Objective To explore the value of a clinical-radiomics-deep learning(CRDL)fusion model based on biparametric mag-netic resonance imaging(bpMRI)in predicting biochemical recurrence(BCR)after radical prostatectomy(RP).Methods A retrospective analysis was conducted on 363 patients with prostate cancer(PCa)confirmed by RP pathology who underwent preoperative MRI,inclu-ding 84 cases experienced BCR(23.1％)and 279 cases did not experience BCR(76.9％).The patients were randomly divided into a training set(n=254)and a test set(n=109)in a ratio of 7∶3.Univariate Cox regression analysis was employed to select clinical variables related to BCR and the clinical model was constructed using backward stepwise multivariate Cox regression analysis.The radiomics features and deep learning(DL)features based on the DenseNet network were extracted.Radiomics and DL signatures were separately developed using least absolute shrinkage and selection operator(LASSO)-Cox regression algorithm.A CRDL fusion model was constructed by combining significant clinical features,DL signature and radiomics signature.The models'predictive performance for BCR was evaluated and compared using the concordance index(C-index).K-M survival curve and Log-rank test were used to assess the performance of CRDL fusion model in risk stratifica-tion of biochemical recurrence free survival(bRFS).Results In the test set,there was no statistically significant difference among C-index of radiomics signature,DL signature and clinical model(P＞0.05).The CRDL fusion model achieved a C-index of 0.83,higher than the clinical model,radiomics signature,and DL signature(P=0.03,0.01,and 0.03).K-M survival curve showed a significant difference in bRFS between low-risk and high-risk patients stratified by the CRDL fusion model[P＜0.000 1,hazard ratio(HR)=30.56,95％confidence interval(CI)10.64-87.75].Conclusion Radiomics signature and DL signature have comparable predictive per-formance for BCR after RP.The CRDL fusion model exhibits the best predictive efficacy for BCR,which is valuable for guiding postoperative treatment strategies in clinical practice.