Sharing standardized image-derived data in computational pathology using DICOM
本文解决了病理图像衍生数据共享不足的问题，通过将这些数据转换为DICOM标准格式并在NCI IDC平台上公开分享来解决。
0 citationsRead paper
Selected work
Representative Papers
CoM$^3$eT: A foundation model for medical image analysis through federated, multidimensional context integration
This study addresses the fragmentation of specialties and limited modality support in existing medical foundation models by proposing a unified visual foundation model integrating pathology and radiology. Leveraging a multi-dimensional context attention mechanism, the framework unifies sparse and dense prediction tasks across 2D and high-dimensional inputs. Combined with multi-task joint training and parameter-efficient fine-tuning (PEFT), it enables federated learning on consumer-grade hardware. This work represents the first cross-specialty, multi-dimensional unified modeling approach for medical vision, achieving state-of-the-art performance across 12 benchmark datasets. Notably, fine-tuning less than 2.5% of parameters yields results comparable to full fine-tuning, while federated learning performance closely approximates centralized training, significantly enhancing adaptation efficiency in low-resource settings.
0 citationsRead paper
Convergent Evolution in Neural Representation Space: Emergent Order in Deep Belief Networks
This study investigates whether deep belief networks (DBNs) trained entirely in an unsupervised manner can spontaneously develop internal representational structures aligned with data categories. Through comprehensive analysis of layer-wise representations on datasets such as MNIST—employing generalized discriminative values (GDV), supervised probing, reconstruction abstraction metrics, effective dimensionality estimation, and free sample generation—the work systematically demonstrates that deeper layers exhibit emergent clustering by class, progressive prototypicality, and markedly enhanced class separability. This clustering effect arises from the learned feature structure rather than random initialization or trivial transformations, providing the first empirical evidence of self-organized emergence of categorical structure in unsupervised deep models.
0 citationsRead paper
Improving Deep Learning-Based Target Volume Auto-Delineation for Adaptive MR-Guided Radiotherapy in Head and Neck Cancer: Impact of a Volume-Aware Dice Loss
This study addresses the challenges of time-consuming manual delineation and frequent omission of small-volume metastatic lymph nodes in head and neck cancer radiotherapy. To overcome these limitations, the authors propose a volume-aware Dice loss function integrated with selective and dual-mask strategies within an nnU-Net ResEnc M architecture for multi-label automatic segmentation of primary tumors and metastatic lymph nodes. By incorporating volume-sensitive weighting, the method enhances detection sensitivity for small lesions. Evaluated on the HNTS-MRG 2024 dataset, the dual-mask approach achieves a lymph node detection sensitivity of 83.46% while maintaining a primary tumor segmentation Dice score of 82.04%. In contrast, the selective strategy improves the lymph node Dice coefficient to 0.758 but at the cost of reduced primary tumor segmentation accuracy.
0 citationsRead paper
Improving Generalization of Deep Learning for Brain Metastases Segmentation Across Institutions
This study addresses the limited generalizability of deep learning models in multi-center brain metastasis segmentation, which arises from variations in imaging devices, acquisition protocols, and patient populations. To mitigate this issue, the authors propose a VAE-MMD preprocessing pipeline that integrates a variational autoencoder with a maximum mean discrepancy (MMD) loss, enhanced by skip connections and self-attention mechanisms. This approach effectively aligns feature distributions across centers without requiring target-domain labels and is combined with nnU-Net to achieve high-precision segmentation. The method substantially improves generalization performance: domain classifier accuracy drops to 0.50, average F1 score increases by 11.1%, symmetric Dice (sDice) improves by 7.93%, Hausdorff distance at the 95th percentile (HD95) decreases by 65.5%, and reconstruction peak signal-to-noise ratio (PSNR) exceeds 36 dB.
0 citationsRead paper
Recent publications
Latest Papers
Sharing standardized image-derived data in computational pathology using DICOM
本文解决了病理图像衍生数据共享不足的问题，通过将这些数据转换为DICOM标准格式并在NCI IDC平台上公开分享来解决。
0 citationsRead paper
CoM$^3$eT: A foundation model for medical image analysis through federated, multidimensional context integration
This study addresses the fragmentation of specialties and limited modality support in existing medical foundation models by proposing a unified visual foundation model integrating pathology and radiology. Leveraging a multi-dimensional context attention mechanism, the framework unifies sparse and dense prediction tasks across 2D and high-dimensional inputs. Combined with multi-task joint training and parameter-efficient fine-tuning (PEFT), it enables federated learning on consumer-grade hardware. This work represents the first cross-specialty, multi-dimensional unified modeling approach for medical vision, achieving state-of-the-art performance across 12 benchmark datasets. Notably, fine-tuning less than 2.5% of parameters yields results comparable to full fine-tuning, while federated learning performance closely approximates centralized training, significantly enhancing adaptation efficiency in low-resource settings.
0 citationsRead paper
Convergent Evolution in Neural Representation Space: Emergent Order in Deep Belief Networks
This study investigates whether deep belief networks (DBNs) trained entirely in an unsupervised manner can spontaneously develop internal representational structures aligned with data categories. Through comprehensive analysis of layer-wise representations on datasets such as MNIST—employing generalized discriminative values (GDV), supervised probing, reconstruction abstraction metrics, effective dimensionality estimation, and free sample generation—the work systematically demonstrates that deeper layers exhibit emergent clustering by class, progressive prototypicality, and markedly enhanced class separability. This clustering effect arises from the learned feature structure rather than random initialization or trivial transformations, providing the first empirical evidence of self-organized emergence of categorical structure in unsupervised deep models.
0 citationsRead paper
Improving Deep Learning-Based Target Volume Auto-Delineation for Adaptive MR-Guided Radiotherapy in Head and Neck Cancer: Impact of a Volume-Aware Dice Loss
This study addresses the challenges of time-consuming manual delineation and frequent omission of small-volume metastatic lymph nodes in head and neck cancer radiotherapy. To overcome these limitations, the authors propose a volume-aware Dice loss function integrated with selective and dual-mask strategies within an nnU-Net ResEnc M architecture for multi-label automatic segmentation of primary tumors and metastatic lymph nodes. By incorporating volume-sensitive weighting, the method enhances detection sensitivity for small lesions. Evaluated on the HNTS-MRG 2024 dataset, the dual-mask approach achieves a lymph node detection sensitivity of 83.46% while maintaining a primary tumor segmentation Dice score of 82.04%. In contrast, the selective strategy improves the lymph node Dice coefficient to 0.758 but at the cost of reduced primary tumor segmentation accuracy.
0 citationsRead paper
Improving Generalization of Deep Learning for Brain Metastases Segmentation Across Institutions
This study addresses the limited generalizability of deep learning models in multi-center brain metastasis segmentation, which arises from variations in imaging devices, acquisition protocols, and patient populations. To mitigate this issue, the authors propose a VAE-MMD preprocessing pipeline that integrates a variational autoencoder with a maximum mean discrepancy (MMD) loss, enhanced by skip connections and self-attention mechanisms. This approach effectively aligns feature distributions across centers without requiring target-domain labels and is combined with nnU-Net to achieve high-precision segmentation. The method substantially improves generalization performance: domain classifier accuracy drops to 0.50, average F1 score increases by 11.1%, symmetric Dice (sDice) improves by 7.93%, Hausdorff distance at the 95th percentile (HD95) decreases by 65.5%, and reconstruction peak signal-to-noise ratio (PSNR) exceeds 36 dB.
0 citationsRead paper