Despite the remarkable successes of transfer learning in materials science,the practicality of existing transfer learning methods are still limited in real-world applications of materials science because they essentia...Despite the remarkable successes of transfer learning in materials science,the practicality of existing transfer learning methods are still limited in real-world applications of materials science because they essentially assume the same material descriptors on source and target materials datasets.In other words,existing transfer learning methods cannot utilize the knowledge extracted from calculated crystal structures when analyzing experimental observations of real-world chemical experiments.We propose a transfer learning criterion,called cross-modality material embedding loss(CroMEL),to build a source feature extractor that can transfer knowledge extracted from calculated crystal structures to prediction models in target applications where only simple chemical compositions are accessible.The prediction models based on transfer learning with CroMEL showed state-of-the-art prediction accuracy on 14 experimental materials datasets in various chemical applications.In particular,the prediction models with CroMEL achieved R2-scores greater than 0.95 in predicting the experimentally measured formation enthalpies and band gaps of the experimentally synthesized materials.展开更多
Background:Source-free unsupervised domain adaptation(SFUDA)methods aim to address the challenge of domain shift while preserving data privacy.Existing SFUDA approaches construct reliable and confident pseudo-labels f...Background:Source-free unsupervised domain adaptation(SFUDA)methods aim to address the challenge of domain shift while preserving data privacy.Existing SFUDA approaches construct reliable and confident pseudo-labels for target-domain data through denoising methods,thereby guiding the training of the target-domain model.The effectiveness of denoising approaches is influenced by the degree of domain gap between the source and target domains.A marked shift can cause the pseudo-labels to be unreliable,even after applying denoising.Methods:We propose a novel 2-stage framework for SFUDA called visual prompt source-free domain adaptation(VP-SFDA).We propose input-specific visual prompt in the first stage,prompting process,which bridges the target-domain data to source-domain distribution.Our method utilizes visual prompts and batch normalization constraint to enable the alignment model to learn domainspecific knowledge and align the target-domain data with the source-domain contribution.The second stage is the adaptation process,which aims at optimizing the segmentation model from the source domain to the target domain.This is accomplished through the denoising techniques,ultimately enhancing the performance.Results:Our study presents a comparative analysis of several SFUDA techniques in the VPSFDA framework across 4 tasks:abdominal magnetic resonance imaging(MRI)to computed tomography(CT),abdominal CT to MRI,cardiac MRI to CT,and cardiac CT to MRI.Notably,in the abdominal MRI to CT adaptation task,the VP-OS method achieved a remarkable improvement,increasing the average DICE score from 0.658 to 0.773(P<0.01)and reducing the average surface distance(ASD)from 3.489 to 2.961(P<0.01).Similarly,the VP-LD and VP-DPL methods also showed significant improvements over their base algorithms in both abdominal and cardiac MRI to CT tasks.Conclusions:This paper proposes VP-SFDA,a novel 2-stage framework for SFUDA in medical imaging,which achieves superior performance through input-specific visual prompts and batch normalization constraint for domain adaptation,coupled with denoising methods for enhanced results.Comparative experiments on 4 medical SFUDA tasks demonstrate that VO-SFDA surpasses existing methods,with ablation studies confirming the benefits of domain-specific patterns.展开更多
基金supported in part by the National Research Foundation(NRF)grant funded by the Korea government(MSIT)(RS-2023-00283597).
文摘Despite the remarkable successes of transfer learning in materials science,the practicality of existing transfer learning methods are still limited in real-world applications of materials science because they essentially assume the same material descriptors on source and target materials datasets.In other words,existing transfer learning methods cannot utilize the knowledge extracted from calculated crystal structures when analyzing experimental observations of real-world chemical experiments.We propose a transfer learning criterion,called cross-modality material embedding loss(CroMEL),to build a source feature extractor that can transfer knowledge extracted from calculated crystal structures to prediction models in target applications where only simple chemical compositions are accessible.The prediction models based on transfer learning with CroMEL showed state-of-the-art prediction accuracy on 14 experimental materials datasets in various chemical applications.In particular,the prediction models with CroMEL achieved R2-scores greater than 0.95 in predicting the experimentally measured formation enthalpies and band gaps of the experimentally synthesized materials.
基金supportted by the Natural Science Foundation of China(62394311,62394310)Beijing Natural Science Foundation(QY24034)National Biomedical Imaging Facility Grant and from the startup funds of Peking University Health Science Center.
文摘Background:Source-free unsupervised domain adaptation(SFUDA)methods aim to address the challenge of domain shift while preserving data privacy.Existing SFUDA approaches construct reliable and confident pseudo-labels for target-domain data through denoising methods,thereby guiding the training of the target-domain model.The effectiveness of denoising approaches is influenced by the degree of domain gap between the source and target domains.A marked shift can cause the pseudo-labels to be unreliable,even after applying denoising.Methods:We propose a novel 2-stage framework for SFUDA called visual prompt source-free domain adaptation(VP-SFDA).We propose input-specific visual prompt in the first stage,prompting process,which bridges the target-domain data to source-domain distribution.Our method utilizes visual prompts and batch normalization constraint to enable the alignment model to learn domainspecific knowledge and align the target-domain data with the source-domain contribution.The second stage is the adaptation process,which aims at optimizing the segmentation model from the source domain to the target domain.This is accomplished through the denoising techniques,ultimately enhancing the performance.Results:Our study presents a comparative analysis of several SFUDA techniques in the VPSFDA framework across 4 tasks:abdominal magnetic resonance imaging(MRI)to computed tomography(CT),abdominal CT to MRI,cardiac MRI to CT,and cardiac CT to MRI.Notably,in the abdominal MRI to CT adaptation task,the VP-OS method achieved a remarkable improvement,increasing the average DICE score from 0.658 to 0.773(P<0.01)and reducing the average surface distance(ASD)from 3.489 to 2.961(P<0.01).Similarly,the VP-LD and VP-DPL methods also showed significant improvements over their base algorithms in both abdominal and cardiac MRI to CT tasks.Conclusions:This paper proposes VP-SFDA,a novel 2-stage framework for SFUDA in medical imaging,which achieves superior performance through input-specific visual prompts and batch normalization constraint for domain adaptation,coupled with denoising methods for enhanced results.Comparative experiments on 4 medical SFUDA tasks demonstrate that VO-SFDA surpasses existing methods,with ablation studies confirming the benefits of domain-specific patterns.
