Pretrained, Curriculum-Tuned, and Ensembled: A Tracer-Aware Interactive Segmentation Pipeline for AutoPET V
本文提出TRIAGE方法，通过预训练、课程调优和集成学习解决全身PET/CT中交互式病灶分割问题。
0 citationsRead paper
Selected work
Representative Papers
MIRAGE: Multi-scale Lesion-Informed Representation with Auxiliary Guidance for MRI Contrast Enhancement
This work addresses the challenge of synthesizing contrast-enhanced breast MRI from a single non-contrast slice, where uncertainty in lesion enhancement complicates the simultaneous preservation of global realism and lesion fidelity. To this end, the authors propose MIRAGE, a method built upon a residual 2D U-Net architecture that integrates global reconstruction and perceptual losses, and introduces three lesion-aware supervision mechanisms tailored to distinct training phases: asymmetric tumor enhancement deficiency penalty, multi-scale auxiliary tumor segmentation, and frozen nnU-Net guidance. Evaluated on a multi-center MAMA-SYNTH dataset of 301 cases, MIRAGE outperforms baseline approaches—including pix2pix and conditional diffusion models—on six out of eight quantitative metrics and demonstrates superior performance in downstream lesion localization tasks.
0 citationsRead paper
Cross-Attention Multimodal Learning for Predicting Response to Neoadjuvant Imatinib in Gastrointestinal Stromal Tumors: A Multicenter Retrospective Study
Current clinical and molecular biomarkers lack reliable predictive power for gastrointestinal stromal tumor (GIST) patient response to neoadjuvant imatinib. To address this challenge, this study proposes an interpretable multimodal deep learning framework that integrates CT imaging with clinical variables through a cross-attention mechanism to enable synergistic cross-modal modeling. The approach leverages self-supervised pretraining, low-rank adaptation, and SMAC3-based hyperparameter optimization. The model achieves an AUC of 0.99 in internal validation, though external test performance ranges from 0.60 to 0.63. Interpretability analyses identify key discriminative features—including CD117, BRAF, PDGFRA, age, and sex—with FDR-corrected P-values ≤ 0.036, offering novel insights for personalized therapeutic decision-making.
0 citationsRead paper
SAFE-Diff: Scale-Aware Attention and Feature-Dispersive Diffusion with Uncertainty Estimation for Contrast-Enhanced Breast MRI Synthesis
Synthesizing high-fidelity contrast-enhanced breast MRI images remains challenging due to the complex texture of lesions and highly heterogeneous enhancement patterns, which hinder the accuracy and efficiency of breast cancer screening. This work proposes a novel diffusion-based generative model that innovatively integrates a scale-aware attention mechanism to capture multi-scale lesion characteristics, employs a feature-dispersed diffusion strategy to enhance textural diversity, and incorporates Bayesian uncertainty estimation to improve clinical reliability. The method significantly outperforms existing approaches in terms of image fidelity, lesion structural preservation, and consistency of enhancement patterns, thereby substantially enhancing the clinical utility of synthesized images.
0 citationsRead paper
LoGo-MR: Screening Breast MRI for Cancer Risk Prediction by Efficient Omni-Slice Modeling
Current breast MRI methods struggle to balance computational efficiency with inter-slice continuity and lack effective stratification of short- to long-term (1–5 years) breast cancer risk. To address this, this work proposes the LoGo-MR framework, which captures local subtle features through adjacent slice encoding for short-term risk prediction and employs Transformer-enhanced multiple instance learning to model global distribution patterns for long-term risk assessment. The framework further integrates axial, sagittal, and coronal multiplanar inputs to generate voxel-level interpretable risk saliency maps. Evaluated on a cohort of approximately 7.5K subjects, LoGo-MR achieves AUCs of 0.77–0.69 for 1–5 year risk prediction and improves the C-index by approximately 6% over 3D CNN baselines. Its multiplanar extension, LoGo3-MR, yields further performance gains and enables cross-planar lesion localization.
0 citationsRead paper
Recent publications
Latest Papers
Pretrained, Curriculum-Tuned, and Ensembled: A Tracer-Aware Interactive Segmentation Pipeline for AutoPET V
本文提出TRIAGE方法，通过预训练、课程调优和集成学习解决全身PET/CT中交互式病灶分割问题。
0 citationsRead paper
MIRAGE: Multi-scale Lesion-Informed Representation with Auxiliary Guidance for MRI Contrast Enhancement
This work addresses the challenge of synthesizing contrast-enhanced breast MRI from a single non-contrast slice, where uncertainty in lesion enhancement complicates the simultaneous preservation of global realism and lesion fidelity. To this end, the authors propose MIRAGE, a method built upon a residual 2D U-Net architecture that integrates global reconstruction and perceptual losses, and introduces three lesion-aware supervision mechanisms tailored to distinct training phases: asymmetric tumor enhancement deficiency penalty, multi-scale auxiliary tumor segmentation, and frozen nnU-Net guidance. Evaluated on a multi-center MAMA-SYNTH dataset of 301 cases, MIRAGE outperforms baseline approaches—including pix2pix and conditional diffusion models—on six out of eight quantitative metrics and demonstrates superior performance in downstream lesion localization tasks.
0 citationsRead paper
Cross-Attention Multimodal Learning for Predicting Response to Neoadjuvant Imatinib in Gastrointestinal Stromal Tumors: A Multicenter Retrospective Study
Current clinical and molecular biomarkers lack reliable predictive power for gastrointestinal stromal tumor (GIST) patient response to neoadjuvant imatinib. To address this challenge, this study proposes an interpretable multimodal deep learning framework that integrates CT imaging with clinical variables through a cross-attention mechanism to enable synergistic cross-modal modeling. The approach leverages self-supervised pretraining, low-rank adaptation, and SMAC3-based hyperparameter optimization. The model achieves an AUC of 0.99 in internal validation, though external test performance ranges from 0.60 to 0.63. Interpretability analyses identify key discriminative features—including CD117, BRAF, PDGFRA, age, and sex—with FDR-corrected P-values ≤ 0.036, offering novel insights for personalized therapeutic decision-making.
0 citationsRead paper
SAFE-Diff: Scale-Aware Attention and Feature-Dispersive Diffusion with Uncertainty Estimation for Contrast-Enhanced Breast MRI Synthesis
Synthesizing high-fidelity contrast-enhanced breast MRI images remains challenging due to the complex texture of lesions and highly heterogeneous enhancement patterns, which hinder the accuracy and efficiency of breast cancer screening. This work proposes a novel diffusion-based generative model that innovatively integrates a scale-aware attention mechanism to capture multi-scale lesion characteristics, employs a feature-dispersed diffusion strategy to enhance textural diversity, and incorporates Bayesian uncertainty estimation to improve clinical reliability. The method significantly outperforms existing approaches in terms of image fidelity, lesion structural preservation, and consistency of enhancement patterns, thereby substantially enhancing the clinical utility of synthesized images.
0 citationsRead paper
LoGo-MR: Screening Breast MRI for Cancer Risk Prediction by Efficient Omni-Slice Modeling
Current breast MRI methods struggle to balance computational efficiency with inter-slice continuity and lack effective stratification of short- to long-term (1–5 years) breast cancer risk. To address this, this work proposes the LoGo-MR framework, which captures local subtle features through adjacent slice encoding for short-term risk prediction and employs Transformer-enhanced multiple instance learning to model global distribution patterns for long-term risk assessment. The framework further integrates axial, sagittal, and coronal multiplanar inputs to generate voxel-level interpretable risk saliency maps. Evaluated on a cohort of approximately 7.5K subjects, LoGo-MR achieves AUCs of 0.77–0.69 for 1–5 year risk prediction and improves the C-index by approximately 6% over 3D CNN baselines. Its multiplanar extension, LoGo3-MR, yields further performance gains and enables cross-planar lesion localization.
0 citationsRead paper