Artificial Intelligence in Medicine最新文献_第2页

BRLA-DDI: A novel framework for drug–drug interaction extraction BRLA-DDI：药物-药物相互作用提取的新框架

IF 6.2 2区医学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Artificial Intelligence in Medicine

Pub Date : 2026-01-21 DOI: 10.1016/j.artmed.2026.103353

Zhu Yuan , Shuailiang Zhang , Zongjin Li , Huiyun Zhang , Huaqi Zhang , Yaxun Jia

Drug–drug interaction (DDI) extraction is a pivotal task in biomedical information processing, focused on identifying potentially adverse drug reactions (ADRs). Despite significant progress in DDI extraction, existing models struggle with complex sentence structures and ambiguous interactions, especially in cases involving rare or implicit drug relationships. To overcome these limitations, this paper presents a novel model, BRLA-DDI, that integrates BioBERT-LSTM mechanism, Relational Graph Convolutional Network (R-GCN), and a loss function incorporating attention (loss+attention) to enhance both accuracy and generalization in DDI tasks. The core innovation of BRLA-DDI lies in its synergistic integration of these components, coupled with two unique methodological contributions. First, the model employs BioBERT and BiLSTM for text feature extraction, effectively leveraging the contextual information within drug descriptions. Second, by thoroughly integrating the multihead attention mechanism with R-GCN, BRLA-DDI strengthens its capability to capture intricate relationships between drug entities. Additionally, we introduce an innovative loss-attention mechanism that merges cross-entropy loss with an attention-based regularization term, offering precise guidance for the model in learning key features during the optimization process. Lastly, we employ a dynamic negative sampling strategy that mitigates the zero-loss issue prevalent in traditional methods, thereby accelerating model convergence and enhancing robustness. Experimental results demonstrate the superiority of the proposed BRLA-DDI approach, achieving a precision of 87.68%, a recall of 88.06%, and an F1 Score of 87.87% on the DDI Extraction 2013 dataset, surpassing a wide range of existing methods. Crucially, the model also exhibits robust and superior performance on the external TAC 2018 dataset, providing compelling evidence of its strong generalizability across different data sources and annotation styles. All our code and data have been publicly released at https://github.com/Hero-Legend/LossAtt-DDI.

药物相互作用（DDI）信息提取是生物医学信息处理中的一项关键任务，其重点是识别潜在的药物不良反应（adr）。尽管在DDI提取方面取得了重大进展，但现有的模型难以处理复杂的句子结构和模棱两可的相互作用，特别是在涉及罕见或隐性药物关系的情况下。为了克服这些限制，本文提出了一种新的模型BRLA-DDI，该模型集成了BioBERT-LSTM机制，关系图卷积网络（R-GCN）和包含注意力（loss+attention）的损失函数，以提高DDI任务的准确性和泛化性。BRLA-DDI的核心创新在于其对这些组成部分的协同整合，以及两种独特的方法贡献。首先，该模型采用BioBERT和BiLSTM进行文本特征提取，有效利用了药物描述中的上下文信息。其次，通过将多头注意机制与R-GCN彻底整合，BRLA-DDI增强了其捕捉药物实体之间复杂关系的能力。此外，我们引入了一种创新的损失-注意机制，该机制将交叉熵损失与基于注意的正则化项相结合，为模型在优化过程中学习关键特征提供了精确的指导。最后，我们采用了一种动态负采样策略，减轻了传统方法中普遍存在的零损失问题，从而加速了模型的收敛并增强了鲁棒性。实验结果表明，BRLA-DDI方法在DDI Extraction 2013数据集上的准确率为87.68%，召回率为88.06%，F1分数为87.87%，超过了现有的许多方法。至关重要的是，该模型在外部TAC 2018数据集上也表现出鲁棒性和卓越的性能，提供了令人信服的证据，证明其在不同数据源和注释样式之间具有强大的泛化能力。我们所有的代码和数据都在https://github.com/Hero-Legend/LossAtt-DDI上公开发布。

{"title":"BRLA-DDI: A novel framework for drug–drug interaction extraction","authors":"Zhu Yuan , Shuailiang Zhang , Zongjin Li , Huiyun Zhang , Huaqi Zhang , Yaxun Jia","doi":"10.1016/j.artmed.2026.103353","DOIUrl":"10.1016/j.artmed.2026.103353","url":null,"abstract":"<div><div>Drug–drug interaction (DDI) extraction is a pivotal task in biomedical information processing, focused on identifying potentially adverse drug reactions (ADRs). Despite significant progress in DDI extraction, existing models struggle with complex sentence structures and ambiguous interactions, especially in cases involving rare or implicit drug relationships. To overcome these limitations, this paper presents a novel model, BRLA-DDI, that integrates BioBERT-LSTM mechanism, Relational Graph Convolutional Network (R-GCN), and a loss function incorporating attention (loss+attention) to enhance both accuracy and generalization in DDI tasks. The core innovation of BRLA-DDI lies in its synergistic integration of these components, coupled with two unique methodological contributions. First, the model employs BioBERT and BiLSTM for text feature extraction, effectively leveraging the contextual information within drug descriptions. Second, by thoroughly integrating the multihead attention mechanism with R-GCN, BRLA-DDI strengthens its capability to capture intricate relationships between drug entities. Additionally, we introduce an innovative loss-attention mechanism that merges cross-entropy loss with an attention-based regularization term, offering precise guidance for the model in learning key features during the optimization process. Lastly, we employ a dynamic negative sampling strategy that mitigates the zero-loss issue prevalent in traditional methods, thereby accelerating model convergence and enhancing robustness. Experimental results demonstrate the superiority of the proposed BRLA-DDI approach, achieving a precision of 87.68%, a recall of 88.06%, and an F1 Score of 87.87% on the DDI Extraction 2013 dataset, surpassing a wide range of existing methods. Crucially, the model also exhibits robust and superior performance on the external TAC 2018 dataset, providing compelling evidence of its strong generalizability across different data sources and annotation styles. All our code and data have been publicly released at <span><span>https://github.com/Hero-Legend/LossAtt-DDI</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55458,"journal":{"name":"Artificial Intelligence in Medicine","volume":"174 ","pages":"Article 103353"},"PeriodicalIF":6.2,"publicationDate":"2026-01-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039453","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Rethinking U-Net architecture in medical imaging: Advancing the efficient and interpretable UKAN-CBAM framework for colorectal polyp segmentation 重新思考医学成像中的U-Net架构：推进结肠直肠息肉分割的高效和可解释的UKAN-CBAM框架。

IF 6.2 2区医学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Artificial Intelligence in Medicine

Pub Date : 2026-01-15 DOI: 10.1016/j.artmed.2026.103352

Md. Faysal Ahamed , Fariya Bintay Shafi , Md. Rabiul Islam , Md. Fahmidun Nabi , Julfikar Haider

Prompt detection of colorectal polyps is essential for preventing colorectal cancer, a leading cause of cancer-related deaths worldwide. However, manual detection through medical imaging faces significant challenges, including high costs, reliance on skilled endoscopists, and susceptibility to errors, which can result in missed diagnoses and adverse health outcomes. This study proposes UKAN-CBAM, an advanced semantic segmentation framework that combines Kolmogorov-Arnold Networks (KANs) with Convolutional Block Attention Modules (CBAM) within a U-Net architecture. This two-phase encoder-decoder design integrates convolutional and tokenized KAN blocks to leverage the efficiency of KANs and the feature refinement capabilities of CBAM, achieving superior segmentation performance with enhanced interpretability and compactness. The framework was trained on the Kvasir-SEG dataset and validated across external datasets, including CVC-ClinicDB, CVC-ColonDB, EndoScene, PolypGen, ETIS-LaribPolypDB, and Piccolo. In addition, 10-fold cross-validation was performed to ensure robustness and generalization. UKAN-CBAM outperformed state-of-the-art (SOTA) methods, achieving an mDice of 93.80%, an mIoU of 89.18%, a precision of 95.65%, a recall of 92.02%, and an accuracy of 96.21%. It also demonstrated computational efficiency, requiring only 55.99 MB of memory and 5.214 GFLOPs, and achieved inference speeds of 122.272 ms per prediction. The feature maps, heatmaps, and Grad-CAM showed that the model focuses on key regions, whereas the ablations highlight the importance of configuration for robustness. Paired t-tests with P values, confidence intervals, and standard deviations, along with 10-fold cross-validation, further confirmed that the reported improvements were statistically significant and not due to chance. Strong generalization across diverse image and video datasets and real-time capabilities provide an effective and reliable tool for clinical applications. This integration of attention mechanisms and interpretability represents a significant step forward in medical diagnostics. Code availability: https://github.com/Faysal425/UKAN_CBAM_Segmentation

结肠直肠息肉的及时发现对于预防结直肠癌至关重要，结直肠癌是全球癌症相关死亡的主要原因。然而，通过医学成像进行人工检测面临着重大挑战，包括成本高、依赖熟练的内窥镜医生以及容易出错，这可能导致漏诊和不良的健康结果。本研究提出了UKAN-CBAM，这是一种先进的语义分割框架，将U-Net架构中的Kolmogorov-Arnold网络（KANs）与卷积块注意模块（CBAM）相结合。这种两阶段编码器-解码器设计集成了卷积和标记化的KAN块，以利用KAN的效率和CBAM的特征细化能力，通过增强的可解释性和紧凑性实现卓越的分割性能。该框架在Kvasir-SEG数据集上进行了训练，并在外部数据集上进行了验证，包括CVC-ClinicDB、CVC-ColonDB、EndoScene、polygen、ETIS-LaribPolypDB和Piccolo。此外，还进行了10次交叉验证，以确保稳健性和泛化。UKAN-CBAM优于最先进的SOTA方法，mDice为93.80%，mIoU为89.18%，精密度为95.65%，召回率为92.02%，准确率为96.21%。它还证明了计算效率，只需要55.99 MB内存和5.214 GFLOPs，每次预测的推理速度达到122.272 ms。特征图、热图和Grad-CAM表明该模型专注于关键区域，而消融则突出了配置对鲁棒性的重要性。配对t检验，包括P值、置信区间和标准差，以及10倍交叉验证，进一步证实了报告的改善在统计学上是显著的，而不是偶然的。对不同图像和视频数据集的强泛化和实时能力为临床应用提供了有效可靠的工具。这种注意力机制和可解释性的整合代表了医学诊断向前迈出的重要一步。代码可用性：https://github.com/Faysal425/UKAN_CBAM_Segmentation。

{"title":"Rethinking U-Net architecture in medical imaging: Advancing the efficient and interpretable UKAN-CBAM framework for colorectal polyp segmentation","authors":"Md. Faysal Ahamed , Fariya Bintay Shafi , Md. Rabiul Islam , Md. Fahmidun Nabi , Julfikar Haider","doi":"10.1016/j.artmed.2026.103352","DOIUrl":"10.1016/j.artmed.2026.103352","url":null,"abstract":"<div><div>Prompt detection of colorectal polyps is essential for preventing colorectal cancer, a leading cause of cancer-related deaths worldwide. However, manual detection through medical imaging faces significant challenges, including high costs, reliance on skilled endoscopists, and susceptibility to errors, which can result in missed diagnoses and adverse health outcomes. This study proposes UKAN-CBAM, an advanced semantic segmentation framework that combines Kolmogorov-Arnold Networks (KANs) with Convolutional Block Attention Modules (CBAM) within a U-Net architecture. This two-phase encoder-decoder design integrates convolutional and tokenized KAN blocks to leverage the efficiency of KANs and the feature refinement capabilities of CBAM, achieving superior segmentation performance with enhanced interpretability and compactness. The framework was trained on the Kvasir-SEG dataset and validated across external datasets, including CVC-ClinicDB, CVC-ColonDB, EndoScene, PolypGen, ETIS-LaribPolypDB, and Piccolo. In addition, 10-fold cross-validation was performed to ensure robustness and generalization. UKAN-CBAM outperformed state-of-the-art (SOTA) methods, achieving an mDice of 93.80%, an mIoU of 89.18%, a precision of 95.65%, a recall of 92.02%, and an accuracy of 96.21%. It also demonstrated computational efficiency, requiring only 55.99 MB of memory and 5.214 GFLOPs, and achieved inference speeds of 122.272 ms per prediction. The feature maps, heatmaps, and Grad-CAM showed that the model focuses on key regions, whereas the ablations highlight the importance of configuration for robustness. Paired <em>t-</em>tests with <em>P</em> values, confidence intervals, and standard deviations, along with 10-fold cross-validation, further confirmed that the reported improvements were statistically significant and not due to chance. Strong generalization across diverse image and video datasets and real-time capabilities provide an effective and reliable tool for clinical applications. This integration of attention mechanisms and interpretability represents a significant step forward in medical diagnostics. Code availability: <span><span>https://github.com/Faysal425/UKAN_CBAM_Segmentation</span><svg><path></path></svg></span></div></div>","PeriodicalId":55458,"journal":{"name":"Artificial Intelligence in Medicine","volume":"174 ","pages":"Article 103352"},"PeriodicalIF":6.2,"publicationDate":"2026-01-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146013422","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Mitigating data center bias in cancer classification: Transfer bias unlearning and feature size reduction via conflict-of-interest free multi-objective optimization 缓解癌症分类中的数据中心偏差：通过无利益冲突的多目标优化消除迁移偏差和减少特征大小

IF 6.2 2区医学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Artificial Intelligence in Medicine

Pub Date : 2026-01-10 DOI: 10.1016/j.artmed.2026.103351

Farnaz Kheiri , Shahryar Rahnamayan , Masoud Makrehchi

Bias in the decision-making processes of trained deep models poses a significant threat to their reliability. Such bias can lead to overoptimistic results on observed data while compromising generalization to unseen datasets. Training data may contain hidden patterns related to task-irrelevant attributes, such as data centers, causing models to exploit these unintended correlations rather than learning the main task. This results in biased predictions that favor certain attributes. To address this issue, we propose an unlearning approach based on Conflict-of-Interest-Free Multi-Objective Optimization, designed to train an unlearning layer that explicitly reduces reliance on irrelevant patterns. Our method aims to minimize the gap between internal accuracy (evaluated on data centers seen during training) and external accuracy (evaluated on entirely unseen data centers) caused by biased model behavior. As a case study, we investigate how data center-specific signatures embedded in cancerous features can lead to misleadingly high internal performance and a significant drop in performance on test samples from external data centers. By evaluating various methods and objective functions, our proposed approach achieves strong generalizability on external validation data by jointly reducing feature dimensionality and excluding conflict-of-interest samples during the

k

-Nearest Neighbor (KNN) searching process. We compare our method against multi-task and adversarial learning approaches for bias mitigation. Results show that our method outperforms others in narrowing the internal-external performance gap while also improving external validation accuracy. To ensure robustness, we conducted experiments using k-fold cross-validation across k different data centers, further validating the generalizability of our approach. Although this study focuses on cancer-related features and data center biases, the proposed method is model-agnostic and can be applied to any biased feature set extracted by a deep learning model.

经过训练的深度模型决策过程中的偏差对其可靠性构成了严重威胁。这种偏差可能导致对观测数据的过度乐观结果，同时损害对未见数据集的泛化。训练数据可能包含与任务无关的属性（如数据中心）相关的隐藏模式，导致模型利用这些非预期的相关性，而不是学习主要任务。这导致偏向某些属性的有偏见的预测。为了解决这个问题，我们提出了一种基于无利益冲突多目标优化的学习方法，旨在训练一个明确减少对不相关模式依赖的学习层。我们的方法旨在最小化内部精度（在训练期间看到的数据中心上进行评估）和外部精度（在完全看不见的数据中心上进行评估）之间的差距，这是由有偏差的模型行为引起的。作为一个案例研究，我们研究了嵌入在癌变特征中的数据中心特定签名如何导致内部性能高得令人误解，以及外部数据中心测试样本的性能显著下降。通过对各种方法和目标函数的评估，我们提出的方法在k-最近邻（KNN）搜索过程中通过联合降低特征维数和排除利益冲突样本，实现了对外部验证数据的强泛化。我们将我们的方法与多任务和对抗性学习方法进行比较，以减轻偏见。结果表明，我们的方法在缩小内外性能差距的同时也提高了外部验证的准确性。为了确保稳健性，我们在k个不同的数据中心进行了k倍交叉验证实验，进一步验证了我们方法的普遍性。尽管本研究侧重于癌症相关特征和数据中心偏差，但所提出的方法是模型不可知的，可以应用于任何由深度学习模型提取的有偏差的特征集。

{"title":"Mitigating data center bias in cancer classification: Transfer bias unlearning and feature size reduction via conflict-of-interest free multi-objective optimization","authors":"Farnaz Kheiri , Shahryar Rahnamayan , Masoud Makrehchi","doi":"10.1016/j.artmed.2026.103351","DOIUrl":"10.1016/j.artmed.2026.103351","url":null,"abstract":"<div><div>Bias in the decision-making processes of trained deep models poses a significant threat to their reliability. Such bias can lead to overoptimistic results on observed data while compromising generalization to unseen datasets. Training data may contain hidden patterns related to task-irrelevant attributes, such as data centers, causing models to exploit these unintended correlations rather than learning the main task. This results in biased predictions that favor certain attributes. To address this issue, we propose an unlearning approach based on Conflict-of-Interest-Free Multi-Objective Optimization, designed to train an unlearning layer that explicitly reduces reliance on irrelevant patterns. Our method aims to minimize the gap between internal accuracy (evaluated on data centers seen during training) and external accuracy (evaluated on entirely unseen data centers) caused by biased model behavior. As a case study, we investigate how data center-specific signatures embedded in cancerous features can lead to misleadingly high internal performance and a significant drop in performance on test samples from external data centers. By evaluating various methods and objective functions, our proposed approach achieves strong generalizability on external validation data by jointly reducing feature dimensionality and excluding conflict-of-interest samples during the <span><math><mi>k</mi></math></span>-Nearest Neighbor (KNN) searching process. We compare our method against multi-task and adversarial learning approaches for bias mitigation. Results show that our method outperforms others in narrowing the internal-external performance gap while also improving external validation accuracy. To ensure robustness, we conducted experiments using k-fold cross-validation across k different data centers, further validating the generalizability of our approach. Although this study focuses on cancer-related features and data center biases, the proposed method is model-agnostic and can be applied to any biased feature set extracted by a deep learning model.</div></div>","PeriodicalId":55458,"journal":{"name":"Artificial Intelligence in Medicine","volume":"174 ","pages":"Article 103351"},"PeriodicalIF":6.2,"publicationDate":"2026-01-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145993609","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Siamese evolutionary masking: Enhancing the generalization of self-supervised medical image segmentation model 暹罗进化掩蔽：增强自监督医学图像分割模型的泛化。

IF 6.2 2区医学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Artificial Intelligence in Medicine

Pub Date : 2026-01-07 DOI: 10.1016/j.artmed.2026.103349

Yichen Zhi, Hongxia Bie, Jiali Wang, Zhao Jing

Self-supervised learning autonomously extracts features from unlabeled data, supporting downstream segmentation tasks with limited annotations. However, variations in devices, imaging parameters, and other factors lead to differences in the distribution of medical images, resulting in poor model generalizability. Mainstream frameworks include Instance Discrimination, which learns features from different perspectives of the same image but may miss details, and Masked Image Modeling (MIM), which captures local features by predicting masked areas but lacks global information capture. To enhance generalizability by combining global and local information, We introduce the Siamese Evolutionary Masking (SEM) framework, which employs a Siamese architecture composed of an online branch and a target branch. An evolutionary masking strategy is adopted within the online branch, transitioning from grid to block masking during training, encouraging the model to develop more general visual features. Additionally, a module called Switch Decoder aligns the online branch’s predicted features with the true features in the target branch, overcoming the challenge of balancing global and local information capture. Experiments on six public datasets, including four skin datasets (SD-260, ISIC2019, ISIC2017, and PH

^{2}

) and two chest X-ray datasets (Chest X-ray PD and Chest X-ray), demonstrate that SEM achieves strong performance among self-supervised methods. In cross-dataset experiments with different distributions, SEM demonstrated the best segmentation and generalization performance, with Dice scores of 81.8% and 91.1%, Jaccard indices of 72.2% and 84.4%, and optimal HD95% measurements of 13.1% and 10.5%, respectively. Code is available at https://github.com/wsdl666/Siamese-Evolutionary-Masking.

自监督学习可以自主地从未标记的数据中提取特征，支持具有有限注释的下游分割任务。然而，由于设备、成像参数等因素的不同，导致医学图像分布的差异，导致模型的泛化性较差。主流框架包括Instance Discrimination（从同一图像的不同角度学习特征，但可能会遗漏细节）和mask image Modeling (MIM)（通过预测mask区域捕获局部特征，但缺乏全局信息捕获）。为了通过结合全局和局部信息来增强可泛化性，我们引入了Siamese进化掩蔽（SEM）框架，该框架采用由在线分支和目标分支组成的Siamese架构。在线分支采用进化掩蔽策略，在训练过程中从网格掩蔽过渡到块掩蔽，鼓励模型发展更通用的视觉特征。此外，一个名为Switch Decoder的模块将在线分支的预测特征与目标分支的真实特征对齐，克服了平衡全局和局部信息捕获的挑战。在6个公共数据集上，包括4个皮肤数据集（SD-260、ISIC2019、ISIC2017和PH2）和2个胸部x射线数据集（chest X-ray PD和chest X-ray），实验表明SEM在自监督方法中具有较强的性能。在不同分布的跨数据集实验中，SEM表现出最佳的分割和泛化性能，Dice得分分别为81.8%和91.1%，Jaccard指数分别为72.2%和84.4%，最优HD95%测量值分别为13.1%和10.5%。代码可从https://github.com/wsdl666/Siamese-Evolutionary-Masking获得。

{"title":"Siamese evolutionary masking: Enhancing the generalization of self-supervised medical image segmentation model","authors":"Yichen Zhi, Hongxia Bie, Jiali Wang, Zhao Jing","doi":"10.1016/j.artmed.2026.103349","DOIUrl":"10.1016/j.artmed.2026.103349","url":null,"abstract":"<div><div>Self-supervised learning autonomously extracts features from unlabeled data, supporting downstream segmentation tasks with limited annotations. However, variations in devices, imaging parameters, and other factors lead to differences in the distribution of medical images, resulting in poor model generalizability. Mainstream frameworks include Instance Discrimination, which learns features from different perspectives of the same image but may miss details, and Masked Image Modeling (MIM), which captures local features by predicting masked areas but lacks global information capture. To enhance generalizability by combining global and local information, We introduce the Siamese Evolutionary Masking (SEM) framework, which employs a Siamese architecture composed of an online branch and a target branch. An evolutionary masking strategy is adopted within the online branch, transitioning from grid to block masking during training, encouraging the model to develop more general visual features. Additionally, a module called Switch Decoder aligns the online branch’s predicted features with the true features in the target branch, overcoming the challenge of balancing global and local information capture. Experiments on six public datasets, including four skin datasets (SD-260, ISIC2019, ISIC2017, and PH<span><math><msup><mrow></mrow><mrow><mn>2</mn></mrow></msup></math></span>) and two chest X-ray datasets (Chest X-ray PD and Chest X-ray), demonstrate that SEM achieves strong performance among self-supervised methods. In cross-dataset experiments with different distributions, SEM demonstrated the best segmentation and generalization performance, with Dice scores of 81.8% and 91.1%, Jaccard indices of 72.2% and 84.4%, and optimal HD95% measurements of 13.1% and 10.5%, respectively. Code is available at <span><span>https://github.com/wsdl666/Siamese-Evolutionary-Masking</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55458,"journal":{"name":"Artificial Intelligence in Medicine","volume":"173 ","pages":"Article 103349"},"PeriodicalIF":6.2,"publicationDate":"2026-01-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145967942","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Integrating probabilistic trees and causal networks for clinical and epidemiological data 整合临床和流行病学数据的概率树和因果网络

IF 6.2 2区医学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Artificial Intelligence in Medicine

Pub Date : 2026-01-06 DOI: 10.1016/j.artmed.2026.103350

Sheresh Zahoor , Pietro Liò , Gaël Dias , Mohammed Hasanuzzaman

Healthcare decision-making requires not only accurate predictions but also insights into how factors influence patient outcomes. While traditional machine learning (ML) models excel at predicting outcomes, such as identifying high-risk patients, they are limited in addressing “what if” questions about interventions. This study introduces the Probabilistic Causal Fusion (PCF) framework, which integrates Causal Bayesian Networks (CBNs) and Probability Trees (PTrees) to extend beyond predictions. PCF leverages causal relationships from CBNs to structure PTrees, enabling both the quantification of factor impacts and the simulation of hypothetical interventions. The framework is evaluated on three clinically diverse, real-world datasets, MIMIC-IV, Framingham Heart Study, and BRFSS (Diabetes), demonstrating consistent predictive performance comparable to conventional ML models, while offering enhanced interpretability and causal reasoning capabilities. In contrast to conventional approaches focused solely on prediction, PCF offers a unified framework for prediction, intervention modelling, and counterfactual analysis, forming a holistic toolkit for clinical decision support. To enhance interpretability, PCF incorporates sensitivity analysis and SHapley Additive exPlanations (SHAP). Sensitivity analysis quantifies the influence of causal parameters on outcomes such as Length of Stay (LOS), Coronary Heart Disease (CHD), and Diabetes, while SHAP highlights the importance of individual features in predictive modelling. This dual-layered interpretability offers both macro-level insights into causal pathways and micro-level explanations for individual predictions. By combining causal reasoning with predictive modelling, PCF bridges the gap between clinical intuition and data-driven insights. Its ability to uncover relationships between modifiable factors and simulate hypothetical scenarios provides clinicians with a clearer understanding of causal pathways. This approach supports more informed, evidence-based decision-making, offering a robust framework for addressing complex questions in diverse healthcare settings.

医疗保健决策不仅需要准确的预测，还需要深入了解各种因素如何影响患者的结果。虽然传统的机器学习（ML）模型擅长预测结果，例如识别高风险患者，但它们在解决有关干预措施的“假设”问题方面受到限制。本研究引入了概率因果融合（PCF）框架，该框架集成了因果贝叶斯网络（CBNs）和概率树（PTrees），以扩展到预测之外。PCF利用cbn的因果关系来构建p树，从而实现因子影响的量化和假设干预的模拟。该框架在三个临床多样化的真实世界数据集（MIMIC-IV、Framingham Heart Study和BRFSS (Diabetes)）上进行了评估，显示出与传统ML模型相当的一致的预测性能，同时提供增强的可解释性和因果推理能力。与仅关注预测的传统方法相比，PCF为预测、干预建模和反事实分析提供了统一的框架，形成了临床决策支持的整体工具包。为了提高可解释性，PCF结合了敏感性分析和SHapley加性解释（SHAP）。敏感性分析量化了因果参数对住院时间（LOS）、冠心病（CHD）和糖尿病等结果的影响，而SHAP强调了预测模型中个体特征的重要性。这种双层可解释性既提供了宏观层面对因果途径的洞察，也为个体预测提供了微观层面的解释。通过将因果推理与预测建模相结合，PCF弥合了临床直觉和数据驱动见解之间的差距。它能够揭示可改变因素之间的关系，并模拟假设情景，为临床医生提供了对因果途径的更清晰理解。这种方法支持更明智的、基于证据的决策，为解决不同医疗保健环境中的复杂问题提供了一个强有力的框架。

{"title":"Integrating probabilistic trees and causal networks for clinical and epidemiological data","authors":"Sheresh Zahoor , Pietro Liò , Gaël Dias , Mohammed Hasanuzzaman","doi":"10.1016/j.artmed.2026.103350","DOIUrl":"10.1016/j.artmed.2026.103350","url":null,"abstract":"<div><div>Healthcare decision-making requires not only accurate predictions but also insights into how factors influence patient outcomes. While traditional machine learning (ML) models excel at predicting outcomes, such as identifying high-risk patients, they are limited in addressing “what if” questions about interventions. This study introduces the Probabilistic Causal Fusion (PCF) framework, which integrates Causal Bayesian Networks (CBNs) and Probability Trees (PTrees) to extend beyond predictions. PCF leverages causal relationships from CBNs to structure PTrees, enabling both the quantification of factor impacts and the simulation of hypothetical interventions. The framework is evaluated on three clinically diverse, real-world datasets, MIMIC-IV, Framingham Heart Study, and BRFSS (Diabetes), demonstrating consistent predictive performance comparable to conventional ML models, while offering enhanced interpretability and causal reasoning capabilities. In contrast to conventional approaches focused solely on prediction, PCF offers a unified framework for prediction, intervention modelling, and counterfactual analysis, forming a holistic toolkit for clinical decision support. To enhance interpretability, PCF incorporates sensitivity analysis and SHapley Additive exPlanations (SHAP). Sensitivity analysis quantifies the influence of causal parameters on outcomes such as Length of Stay (LOS), Coronary Heart Disease (CHD), and Diabetes, while SHAP highlights the importance of individual features in predictive modelling. This dual-layered interpretability offers both macro-level insights into causal pathways and micro-level explanations for individual predictions. By combining causal reasoning with predictive modelling, PCF bridges the gap between clinical intuition and data-driven insights. Its ability to uncover relationships between modifiable factors and simulate hypothetical scenarios provides clinicians with a clearer understanding of causal pathways. This approach supports more informed, evidence-based decision-making, offering a robust framework for addressing complex questions in diverse healthcare settings.</div></div>","PeriodicalId":55458,"journal":{"name":"Artificial Intelligence in Medicine","volume":"173 ","pages":"Article 103350"},"PeriodicalIF":6.2,"publicationDate":"2026-01-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928173","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Tackling data scarcity: Synthetic tumour and mask generation to improve image segmentation 解决数据短缺：合成肿瘤和掩码生成以改善图像分割。

IF 6.2 2区医学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Artificial Intelligence in Medicine

Pub Date : 2025-12-31 DOI: 10.1016/j.artmed.2025.103348

Félix Quinton , Benoit Presles , Romain Popoff , François Godard , Olivier Chevallier , Julie Pellegrinelli , Jean-Marc Vrigneaud , Jean-Louis Alberini , Fabrice Meriaudeau

Given the increasing data requirements of deep learning models and the scarcity of medical imaging data, new data augmentation techniques are receiving particular attention. This paper explores the subfield of tumour synthesis within medical image generation, focusing on the development of synthetic tumours in MR images. This study introduces a novel tumour generation method using diffusion models, designed to inpaint visually convincing 3D synthetic liver tumours into real MRI volumes while generating the corresponding masks using simplex deformation. This approach has been employed successfully to inpaint images with

1 000

synthetic tumours. Furthermore, it has shown significant performance improvements when applied in image segmentation tasks. In particular, our method improved the Dice coefficient by 6.7 points on the ATLAS test set without relying on external data. When combined with a pseudo-annotated external dataset, the improvement increased to 10 points. This study not only demonstrates the ability to segment tumours but also paves the way for various synthetic data-based applications in medical imaging.

鉴于深度学习模型对数据的需求不断增加以及医学成像数据的稀缺，新的数据增强技术受到了特别的关注。本文探讨了医学图像生成中肿瘤合成的子领域，重点是MR图像中合成肿瘤的发展。本研究介绍了一种使用扩散模型的新型肿瘤生成方法，旨在将视觉上令人信服的3D合成肝脏肿瘤绘制到真实的MRI体积中，同时使用单纯形变形生成相应的掩模。这种方法已经成功地用于绘制1000个合成肿瘤的图像。此外，当应用于图像分割任务时，它显示出显着的性能改进。特别是，我们的方法在不依赖外部数据的情况下，将ATLAS测试集上的Dice系数提高了6.7点。当与伪注释的外部数据集结合使用时，改进增加到10分。这项研究不仅证明了肿瘤分割的能力，而且为各种基于合成数据的医学成像应用铺平了道路。

{"title":"Tackling data scarcity: Synthetic tumour and mask generation to improve image segmentation","authors":"Félix Quinton , Benoit Presles , Romain Popoff , François Godard , Olivier Chevallier , Julie Pellegrinelli , Jean-Marc Vrigneaud , Jean-Louis Alberini , Fabrice Meriaudeau","doi":"10.1016/j.artmed.2025.103348","DOIUrl":"10.1016/j.artmed.2025.103348","url":null,"abstract":"<div><div>Given the increasing data requirements of deep learning models and the scarcity of medical imaging data, new data augmentation techniques are receiving particular attention. This paper explores the subfield of tumour synthesis within medical image generation, focusing on the development of synthetic tumours in MR images. This study introduces a novel tumour generation method using diffusion models, designed to inpaint visually convincing 3D synthetic liver tumours into real MRI volumes while generating the corresponding masks using simplex deformation. This approach has been employed successfully to inpaint images with <span><math><mrow><mn>1</mn><mspace></mspace><mn>000</mn></mrow></math></span> synthetic tumours. Furthermore, it has shown significant performance improvements when applied in image segmentation tasks. In particular, our method improved the Dice coefficient by 6.7 points on the ATLAS test set without relying on external data. When combined with a pseudo-annotated external dataset, the improvement increased to 10 points. This study not only demonstrates the ability to segment tumours but also paves the way for various synthetic data-based applications in medical imaging.</div></div>","PeriodicalId":55458,"journal":{"name":"Artificial Intelligence in Medicine","volume":"173 ","pages":"Article 103348"},"PeriodicalIF":6.2,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145914200","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Learning Health Systems provide a glide path to safe landing for AI in health 学习型卫生系统为人工智能在卫生领域的安全着陆提供了一条下坡路

IF 6.2 2区医学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Artificial Intelligence in Medicine

Pub Date : 2025-12-31 DOI: 10.1016/j.artmed.2025.103346

Vasa Curcin , Brendan Delaney , Ahmad Alkhatib , Neil Cockburn , Olivia Dann , Olga Kostopoulou , Daniel Leightley , Matthew Maddocks , Sanjay Modgil , Krishnarajah Nirantharakumar , Philip Scott , Ingrid Wolfe , Kelly Zhang , Charles Friedman

Artificial Intelligence (AI) holds significant promise for healthcare but often struggles to transition from development to clinical integration. This paper argues that Learning Health Systems (LHS)—socio-technical ecosystems designed for continuous data-driven improvement—provide a potential “glide path” for safe, sustainable AI deployment. Just as modern aviation depends on instrument landing systems, the safe and effective integration of AI into healthcare requires the socio-technical infrastructure of LHSs, that enable iterative development and monitoring of AI tools, integrating clinical, technical, and ethical considerations through stakeholder collaboration. They address key challenges in AI implementation, including model generalizability, workflow integration, and transparency, by embedding co-creation, real-world evaluation, and continuous learning into care processes. Unlike static deployments, LHSs support the dynamic evolution of AI systems, incorporating feedback and recalibration to mitigate performance drift and bias. Moreover, they embed governance and regulatory functions—clarifying accountability, supporting data and model provenance, and upholding FAIR (Findable, Accessible, Interoperable, Reusable) principles. LHSs also promote “human-in-the-loop” safety through structured studies of human-AI interaction and shared decision-making. The paper outlines practical steps to align AI with LHS frameworks, including investment in data infrastructure, continuous model monitoring, and fostering a learning culture. Embedding AI in LHSs transforms implementation from a one-time event into a sustained, evidence-based learning process that aligns innovation with clinical realities, ultimately advancing patient care, health equity, and system resilience. The arguments build on insights from an international workshop hosted in 2025, offering a strategic vision for the future of AI in healthcare.

人工智能（AI）为医疗保健带来了巨大的希望，但往往难以从开发过渡到临床整合。本文认为，学习卫生系统（LHS）——为持续数据驱动改进而设计的社会技术生态系统——为安全、可持续的人工智能部署提供了一条潜在的“滑行路径”。正如现代航空依赖于仪表着陆系统一样，将人工智能安全有效地整合到医疗保健中需要lhs的社会技术基础设施，从而实现人工智能工具的迭代开发和监控，并通过利益相关者协作整合临床、技术和道德考虑。它们通过在护理过程中嵌入共同创造、现实世界评估和持续学习，解决了人工智能实施中的关键挑战，包括模型通用性、工作流集成和透明度。与静态部署不同，lhs支持人工智能系统的动态发展，结合反馈和重新校准来减轻性能漂移和偏差。此外，它们嵌入了治理和监管功能——澄清责任、支持数据和模型来源，以及维护FAIR（可查找、可访问、可互操作、可重用）原则。lhs还通过对人类与人工智能互动和共享决策的结构化研究，促进了“人在环”的安全性。本文概述了使人工智能与LHS框架保持一致的实际步骤，包括对数据基础设施的投资、持续的模型监测和培养学习文化。在lhs中嵌入人工智能可以将实施从一次性事件转变为持续的、以证据为基础的学习过程，使创新与临床现实保持一致，最终促进患者护理、卫生公平和系统弹性。这些论点建立在2025年举办的一次国际研讨会的见解基础上，该研讨会为人工智能在医疗保健领域的未来提供了战略愿景。

{"title":"Learning Health Systems provide a glide path to safe landing for AI in health","authors":"Vasa Curcin , Brendan Delaney , Ahmad Alkhatib , Neil Cockburn , Olivia Dann , Olga Kostopoulou , Daniel Leightley , Matthew Maddocks , Sanjay Modgil , Krishnarajah Nirantharakumar , Philip Scott , Ingrid Wolfe , Kelly Zhang , Charles Friedman","doi":"10.1016/j.artmed.2025.103346","DOIUrl":"10.1016/j.artmed.2025.103346","url":null,"abstract":"<div><div>Artificial Intelligence (AI) holds significant promise for healthcare but often struggles to transition from development to clinical integration. This paper argues that Learning Health Systems (LHS)—socio-technical ecosystems designed for continuous data-driven improvement—provide a potential “glide path” for safe, sustainable AI deployment. Just as modern aviation depends on instrument landing systems, the safe and effective integration of AI into healthcare requires the socio-technical infrastructure of LHSs, that enable iterative development and monitoring of AI tools, integrating clinical, technical, and ethical considerations through stakeholder collaboration. They address key challenges in AI implementation, including model generalizability, workflow integration, and transparency, by embedding co-creation, real-world evaluation, and continuous learning into care processes. Unlike static deployments, LHSs support the dynamic evolution of AI systems, incorporating feedback and recalibration to mitigate performance drift and bias. Moreover, they embed governance and regulatory functions—clarifying accountability, supporting data and model provenance, and upholding FAIR (Findable, Accessible, Interoperable, Reusable) principles. LHSs also promote “human-in-the-loop” safety through structured studies of human-AI interaction and shared decision-making. The paper outlines practical steps to align AI with LHS frameworks, including investment in data infrastructure, continuous model monitoring, and fostering a learning culture. Embedding AI in LHSs transforms implementation from a one-time event into a sustained, evidence-based learning process that aligns innovation with clinical realities, ultimately advancing patient care, health equity, and system resilience. The arguments build on insights from an international workshop hosted in 2025, offering a strategic vision for the future of AI in healthcare.</div></div>","PeriodicalId":55458,"journal":{"name":"Artificial Intelligence in Medicine","volume":"173 ","pages":"Article 103346"},"PeriodicalIF":6.2,"publicationDate":"2025-12-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928171","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Smiling difficulties in Alzheimer’s disease linked to reduced nucleus accumbens and pallidum brain volume: Deep learning insights 阿尔茨海默病中的微笑困难与伏隔核和苍白球脑容量减少有关：深度学习见解

IF 6.2 2区医学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Artificial Intelligence in Medicine

Pub Date : 2025-12-30 DOI: 10.1016/j.artmed.2025.103347

Tomomichi Iizuka , Yumi Umeda-Kameyama , Makoto Fukasawa , Masahiro Akishita , Masashi Kameyama

Patients tend to lose the ability to smile during the course of dementia. However, such impairments have rarely been reported, likely due to challenges in quantifying facial expressions. However, feature extraction is now automated due to recent developments in deep learning, which is a machine learning method used in artificial intelligence (AI). We used the output of image-classification AI to quantify smiles in participants with Alzheimer’s disease (AD) and with normal cognition (NC). We found that the ability to form a smile upon request is impaired in patients with AD and that it is associated with reduced volumes of the nucleus accumbens and pallidum. Furthermore, smiling faces were classified with higher accuracy than neutral faces in discriminating between AD and NC. A score from neutral face showed significant correlation with cognitive function. These findings generate hypotheses regarding the neural mechanisms underlying impaired facial expressions in dementia.

在痴呆症的过程中，病人往往会失去微笑的能力。然而，这种损伤很少被报道，可能是由于量化面部表情的挑战。然而，由于深度学习的最新发展，特征提取现在是自动化的，深度学习是一种用于人工智能（AI）的机器学习方法。我们使用图像分类AI的输出来量化阿尔茨海默病（AD）和正常认知（NC）参与者的微笑。我们发现，阿尔茨海默病患者应要求而微笑的能力受损，这与伏隔核和苍白球的体积减少有关。此外，在区分AD和NC方面，微笑面孔的分类准确率高于中性面孔。中性面孔得分与认知功能有显著相关性。这些发现产生了关于痴呆患者面部表情受损的神经机制的假设。

引用次数: 0

UniStain: A unified and organ-aware virtual H&E staining framework for label-free autofluorescence images UniStain：一个统一的和器官感知的虚拟H&E染色框架，用于无标签的自体荧光图像

IF 6.2 2区医学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Artificial Intelligence in Medicine

Pub Date : 2025-12-30 DOI: 10.1016/j.artmed.2025.103335

Lulin Shi , Xingzhong Hou , James K.W. Lai , Ivy H.M. Wong , Bingxin Huang , Athena L.Y. Hui , Ronald C.K. Chan , Terence T.W. Wong

While hematoxylin and eosin (H&E) staining remains the gold standard for pathological diagnosis, its chemical-dependent workflow presents significant limitations, such as time-consuming protocols, hazardous reagent disposal and batch-to-batch variability in stain quality. We present UniStain, a breakthrough virtual staining framework that leverages label-free autofluorescence (AF) imaging and prompt-based deep learning to overcome these challenges. Unlike existing single-organ approaches that require multiple specialized models, our architecture enables versatile multi-tissue staining through a single model, significantly reducing computational overhead. The proposed crosspatch self-attention guidance (CPSG) mechanism addresses critical whole-slide image challenges by maintaining style consistency across adjacent patches and eliminating stitching artifacts. To support comprehensive evaluation, we curate and release the first multi-organ AF/H&E dataset with human tissue samples. Additionally, we introduce downstream clinical validation tasks including image retrieval and cancer subtyping analysis, thereby establishing a robust evaluation framework for virtual staining models. Quantitative assessments (image quality metrics, visual Turing tests) and downstream analyses demonstrate UniStain’s superior performance compared to existing image translation methods, achieving state-of-the-art results while eliminating chemical staining requirements. The dataset and code of UniStain can be found at https://github.com/TABLAB-HKUST/UniStain.

虽然苏木精和伊红（H&；E）染色仍然是病理诊断的金标准，但其化学依赖的工作流程存在显着的局限性，例如耗时的协议，危险试剂的处理以及染色质量批次之间的差异。我们提出UniStain，一个突破性的虚拟染色框架，利用无标签的自体荧光（AF）成像和基于提示的深度学习来克服这些挑战。与现有的单器官方法需要多个专门的模型不同，我们的架构可以通过单个模型实现多组织染色，显著降低计算开销。提出的交叉斑块自注意引导（CPSG）机制通过保持相邻斑块之间的样式一致性和消除拼接伪影来解决关键的整张幻灯片图像挑战。为了支持全面的评估，我们策划并发布了第一个人体组织样本的多器官AF/H&；E数据集。此外，我们还引入了下游临床验证任务，包括图像检索和癌症亚型分析，从而为虚拟染色模型建立了一个强大的评估框架。定量评估（图像质量指标，视觉图灵测试）和下游分析表明，与现有的图像翻译方法相比，UniStain的性能优越，实现了最先进的结果，同时消除了化学染色的要求。UniStain的数据集和代码可以在https://github.com/TABLAB-HKUST/UniStain上找到。

{"title":"UniStain: A unified and organ-aware virtual H&E staining framework for label-free autofluorescence images","authors":"Lulin Shi , Xingzhong Hou , James K.W. Lai , Ivy H.M. Wong , Bingxin Huang , Athena L.Y. Hui , Ronald C.K. Chan , Terence T.W. Wong","doi":"10.1016/j.artmed.2025.103335","DOIUrl":"10.1016/j.artmed.2025.103335","url":null,"abstract":"<div><div>While hematoxylin and eosin (H&E) staining remains the gold standard for pathological diagnosis, its chemical-dependent workflow presents significant limitations, such as time-consuming protocols, hazardous reagent disposal and batch-to-batch variability in stain quality. We present UniStain, a breakthrough virtual staining framework that leverages label-free autofluorescence (AF) imaging and prompt-based deep learning to overcome these challenges. Unlike existing single-organ approaches that require multiple specialized models, our architecture enables versatile multi-tissue staining through a single model, significantly reducing computational overhead. The proposed crosspatch self-attention guidance (CPSG) mechanism addresses critical whole-slide image challenges by maintaining style consistency across adjacent patches and eliminating stitching artifacts. To support comprehensive evaluation, we curate and release the first multi-organ AF/H&E dataset with human tissue samples. Additionally, we introduce downstream clinical validation tasks including image retrieval and cancer subtyping analysis, thereby establishing a robust evaluation framework for virtual staining models. Quantitative assessments (image quality metrics, visual Turing tests) and downstream analyses demonstrate UniStain’s superior performance compared to existing image translation methods, achieving state-of-the-art results while eliminating chemical staining requirements. The dataset and code of UniStain can be found at <span><span>https://github.com/TABLAB-HKUST/UniStain</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55458,"journal":{"name":"Artificial Intelligence in Medicine","volume":"173 ","pages":"Article 103335"},"PeriodicalIF":6.2,"publicationDate":"2025-12-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145852435","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

B2E-CDG: Conditional diffusion-based for label-free OCT angiography artifact removal and robust vascular reconstruction B2E-CDG：基于条件扩散的无标记OCT血管造影伪影去除和稳健血管重建

IF 6.2 2区医学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Artificial Intelligence in Medicine

Pub Date : 2025-12-30 DOI: 10.1016/j.artmed.2025.103345

Jing Xu , Suzhong Fu , Jiwei Xing , Linyan Xue , Qingliang Zhao

Optical Coherence Tomography Angiography (OCTA) is a revolutionary technology widely used in the diagnosis and management of fundus, skin and cardiovascular diseases. However, unavoidable movements, such as breathing, often introduce motion artifacts into OCTA images, which can significantly degrade image quality, obscure critical vascular details, and reduce the diagnostic reliability of the modality. Although recent advances in learning-based image inpainting methods for OCTA enface images have made notable progress in artifact removal, these methods typically require large amounts of accurately labeled data and the generation of pseudo stripes to construct paired training datasets. Additionally, the abundant structural information and flow intensity signals available in OCTA B-scans are often under-utilized. Here we proposed a novel method:B-scans to Enface Conditional Diffusion Guidance (B2E-CDG), which translates signal-void B-scans into normal B-scans. Moreover, the normal B-scans were introduced in a connection manner and the specified reference B-scans in a gradient-based manner as style feature guidance within a diffusion model. Importantly, conditional guidance facilitates a more controlled and precise generation process for flow signal recovery in B-scans. Notably, our method eliminates the need for labeled datasets and pseudo stripes, due to the repetitive scanning nature of OCTA inherently provides paired signal-void and normal B- scans. Our results demonstrated that B2E-CDG effectively removes motion artifacts and restores vascular and structural details. The proposed method shows superior performance in vascular recovery and artifact removal metrics, thereby improving the clinical utility and diagnostic reliability of OCTA.

光学相干断层血管造影（OCTA）是一项革命性的技术，广泛应用于眼底、皮肤和心血管疾病的诊断和治疗。然而，不可避免的运动，如呼吸，通常会在OCTA图像中引入运动伪影，这会显著降低图像质量，模糊关键血管细节，降低模态诊断的可靠性。尽管最近基于学习的OCTA表面图像的图像绘制方法在去除伪影方面取得了显著进展，但这些方法通常需要大量准确标记的数据和生成伪条纹来构建成对的训练数据集。此外，OCTA b扫描中丰富的结构信息和血流强度信号往往没有得到充分利用。本文提出了一种新的方法：b扫描到面条件扩散引导（B2E-CDG），它将信号空洞的b扫描转换为正常的b扫描。此外，在扩散模型中，以连接方式引入正常b扫描，并以基于梯度的方式引入指定参考b扫描作为风格特征指导。重要的是，条件引导有助于b扫描中流量信号恢复的更可控和精确的生成过程。值得注意的是，我们的方法消除了对标记数据集和伪条纹的需要，因为OCTA的重复扫描本质上提供了成对的信号空洞和正常B扫描。我们的研究结果表明，B2E-CDG有效地去除运动伪影，恢复血管和结构细节。该方法在血管恢复和伪影去除指标方面表现出优异的性能，从而提高了OCTA的临床实用性和诊断可靠性。

{"title":"B2E-CDG: Conditional diffusion-based for label-free OCT angiography artifact removal and robust vascular reconstruction","authors":"Jing Xu , Suzhong Fu , Jiwei Xing , Linyan Xue , Qingliang Zhao","doi":"10.1016/j.artmed.2025.103345","DOIUrl":"10.1016/j.artmed.2025.103345","url":null,"abstract":"<div><div>Optical Coherence Tomography Angiography (OCTA) is a revolutionary technology widely used in the diagnosis and management of fundus, skin and cardiovascular diseases. However, unavoidable movements, such as breathing, often introduce motion artifacts into OCTA images, which can significantly degrade image quality, obscure critical vascular details, and reduce the diagnostic reliability of the modality. Although recent advances in learning-based image inpainting methods for OCTA enface images have made notable progress in artifact removal, these methods typically require large amounts of accurately labeled data and the generation of pseudo stripes to construct paired training datasets. Additionally, the abundant structural information and flow intensity signals available in OCTA B-scans are often under-utilized. Here we proposed a novel method:B-scans to Enface Conditional Diffusion Guidance (B2E-CDG), which translates signal-void B-scans into normal B-scans. Moreover, the normal B-scans were introduced in a connection manner and the specified reference B-scans in a gradient-based manner as style feature guidance within a diffusion model. Importantly, conditional guidance facilitates a more controlled and precise generation process for flow signal recovery in B-scans. Notably, our method eliminates the need for labeled datasets and pseudo stripes, due to the repetitive scanning nature of OCTA inherently provides paired signal-void and normal B- scans. Our results demonstrated that B2E-CDG effectively removes motion artifacts and restores vascular and structural details. The proposed method shows superior performance in vascular recovery and artifact removal metrics, thereby improving the clinical utility and diagnostic reliability of OCTA.</div></div>","PeriodicalId":55458,"journal":{"name":"Artificial Intelligence in Medicine","volume":"173 ","pages":"Article 103345"},"PeriodicalIF":6.2,"publicationDate":"2025-12-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145928172","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0