MedPCFM-TED: One-Step Point Cloud Flow Matching for Implant Generation via Teacher-Guided Endpoint Distillation
为解决颅骨植入物生成中多次神经函数评估导致的效率问题,提出了一种基于教师指导端点蒸馏的一步法框架TED,提高了生成速度和质量。
为解决颅骨植入物生成中多次神经函数评估导致的效率问题,提出了一种基于教师指导端点蒸馏的一步法框架TED,提高了生成速度和质量。
本文针对乳腺DCE-MRI合成问题,提出MAMA-FLUX.2方法,基于预训练模型和区域训练目标,通过LoRA微调实现高效适应。
This work addresses the challenge that whole-slide images (WSIs) in digital pathology are acquired across continuously varying magnifications, while existing deep learning models are scale-sensitive and struggle to generalize to unseen or misaligned magnification levels. To overcome this limitation, the authors propose Conditional Layer Normalization (CLN), a lightweight mechanism that employs a small MLP to dynamically generate normalization parameters based on the input pixel size. Integrated into standard CNN architectures and trained on image patches sampled across a continuous range of scales, CLN enables a single model to achieve strong generalization across arbitrary magnifications. Notably, this approach is the first to cover a continuous spectrum of magnifications without requiring ensemble models. On the PANDA prostate cancer dataset, it matches or exceeds the performance of dedicated single-magnification models, consistently ranking among the top three across all evaluated magnifications—including unseen ones—while reducing both training and inference costs by 4–5×.
Clinical data are frequently compromised by missing values, leading to machine learning models with unstable features, poor interpretability, and insufficient robustness—limitations that hinder their deployment in high-stakes clinical decision-making. To address this challenge, this work proposes CACTUS, a novel framework that uniquely prioritizes feature stability as a core evaluation criterion. By integrating feature abstraction, interpretable classification, and systematic stability analysis, CACTUS enables trustworthy predictions even with small-scale, heterogeneous, and incomplete clinical datasets. Evaluated on a cohort of 568 hematuria patients, CACTUS achieves competitive or superior predictive performance while substantially enhancing the stability of key features under missing data conditions. Notably, it demonstrates robustness in sex-stratified analyses, thereby improving the model’s clinical credibility and reproducibility.
In federated learning (FL) for medical imaging, non-IID data arising from inter-site MRI intensity normalization discrepancies—coupled with privacy-preserving constraints that limit data sharing—severely hinder model generalizability and performance. Method: This work systematically evaluates the impact of diverse intensity normalization strategies on 3D brain tumor segmentation and proposes a privacy-preserving FL framework tailored to multi-center heterogeneous data. It identifies normalization choice as a primary driver of client-wise distribution shift and introduces a robust FL training strategy adaptive to heterogeneous intensity distributions. Contribution/Results: Without moving raw data from local sites, the proposed method achieves a 92% 3D Dice score on the BraTS test set—matching centralized training performance—and provides the first empirical validation that high-fidelity, privacy-compliant 3D medical image segmentation is feasible under realistic, clinically observed normalization heterogeneity.
为解决颅骨植入物生成中多次神经函数评估导致的效率问题,提出了一种基于教师指导端点蒸馏的一步法框架TED,提高了生成速度和质量。
本文针对乳腺DCE-MRI合成问题,提出MAMA-FLUX.2方法,基于预训练模型和区域训练目标,通过LoRA微调实现高效适应。
This work addresses the challenge that whole-slide images (WSIs) in digital pathology are acquired across continuously varying magnifications, while existing deep learning models are scale-sensitive and struggle to generalize to unseen or misaligned magnification levels. To overcome this limitation, the authors propose Conditional Layer Normalization (CLN), a lightweight mechanism that employs a small MLP to dynamically generate normalization parameters based on the input pixel size. Integrated into standard CNN architectures and trained on image patches sampled across a continuous range of scales, CLN enables a single model to achieve strong generalization across arbitrary magnifications. Notably, this approach is the first to cover a continuous spectrum of magnifications without requiring ensemble models. On the PANDA prostate cancer dataset, it matches or exceeds the performance of dedicated single-magnification models, consistently ranking among the top three across all evaluated magnifications—including unseen ones—while reducing both training and inference costs by 4–5×.
Clinical data are frequently compromised by missing values, leading to machine learning models with unstable features, poor interpretability, and insufficient robustness—limitations that hinder their deployment in high-stakes clinical decision-making. To address this challenge, this work proposes CACTUS, a novel framework that uniquely prioritizes feature stability as a core evaluation criterion. By integrating feature abstraction, interpretable classification, and systematic stability analysis, CACTUS enables trustworthy predictions even with small-scale, heterogeneous, and incomplete clinical datasets. Evaluated on a cohort of 568 hematuria patients, CACTUS achieves competitive or superior predictive performance while substantially enhancing the stability of key features under missing data conditions. Notably, it demonstrates robustness in sex-stratified analyses, thereby improving the model’s clinical credibility and reproducibility.
In federated learning (FL) for medical imaging, non-IID data arising from inter-site MRI intensity normalization discrepancies—coupled with privacy-preserving constraints that limit data sharing—severely hinder model generalizability and performance. Method: This work systematically evaluates the impact of diverse intensity normalization strategies on 3D brain tumor segmentation and proposes a privacy-preserving FL framework tailored to multi-center heterogeneous data. It identifies normalization choice as a primary driver of client-wise distribution shift and introduces a robust FL training strategy adaptive to heterogeneous intensity distributions. Contribution/Results: Without moving raw data from local sites, the proposed method achieves a 92% 3D Dice score on the BraTS test set—matching centralized training performance—and provides the first empirical validation that high-fidelity, privacy-compliant 3D medical image segmentation is feasible under realistic, clinically observed normalization heterogeneity.