We propose UnEBOLT, a unified end-to-end framework for whole-brain EEG-to-fMRI translation. By combining a multi-dimensional EEG encoder with a Gated Adaptive Fusion module and ROI-specific representation learning, UnEBOLT reconstructs fMRI BOLD signals and functional connectivity across all brain regions in a single model, without region-specific training.
arXiv
VLM-Guided Iterative Refinement for Surgical Image Segmentation with Foundation Models
Ange Lou, Yamin Li, Qi Chang, Nan Xi, Luyuan Xie , Zichao Li, and Tianyu Luan
We propose IR-SIS, an iterative refinement system for surgical image segmentation driven by natural language. IR-SIS combines a fine-tuned SAM3 for initial segmentation with a Vision-Language Model that detects errors and generates corrective prompts, enabling clinician-in-the-loop interaction and adaptive refinement without retraining.
2025
ICML 2025
Collaborative Memory: Multi-User Memory Sharing in LLM Agents with Dynamic Access Control
Alireza Rezazadeh , Zichao Li, Ange Lou, Yuying Zhao, Wei Wei, and Yujia Bao
ICML 2025 Multi-Agent Systems Workshop (ICML 2025), 2025
We introduce Collaborative Memory, a framework for multi-user, multi-agent LLM environments with asymmetric, time-evolving access controls encoded as bipartite graphs. The system maintains private and shared memory tiers, with granular read/write policies that enforce user-agent-resource constraints and enable safe, auditable cross-user knowledge sharing.
SPIE:MI 2025
Zero-Shot Surgical Tool Segmentation in Monocular Video Using Segment Anything Model 2
We explore zero-shot surgical tool segmentation using Segment Anything Model 2 (SAM2) on monocular endoscopic video. Without any domain-specific fine-tuning, SAM2’s promptable video segmentation capability is adapted for tracking and segmenting surgical instruments across frames, achieving competitive performance on standard benchmarks.
2024
NeurIPS 2024
NeuroBOLT: Resting-state EEG-to-fMRI Synthesis with Multi-dimensional Feature Mapping
Yamin Li, Ange Lou, Ziyuan Xu, Shengchao Zhang, Shiyu Wang, Dario J. Englot, Soheil Kolouri, Daniel Moyer, Roza G. Bayrak, and Catie Chang
Advances in Neural Information Processing Systems (NeurIPS 2024), 2024
We present NeuroBOLT, a framework for synthesizing resting-state fMRI BOLD signals from EEG recordings using multi-dimensional feature mapping. Our approach leverages the temporal resolution of EEG and the spatial richness of fMRI to bridge neuroscientific modalities, enabling large-scale brain connectivity studies without simultaneous EEG-fMRI acquisition.
ECCV 2024
Divide and Fuse: Body Part Mesh Recovery from Partially Visible Human Images
We propose a divide-and-fuse strategy for 3D human body part mesh recovery from images where parts of the body may be occluded or out-of-frame. By decomposing the body into independent part estimators and then fusing their outputs, our method achieves robust reconstruction under challenging partial-visibility conditions.
CVPR 2024
DaReNeRF: Direction-aware Representation for Dynamic Scenes
DaReNeRF introduces a direction-aware representation for Neural Radiance Fields applied to dynamic scenes. By explicitly modeling motion direction alongside scene geometry, our method improves novel view synthesis quality for complex non-rigid motions while remaining computationally efficient.
SPIE:MI 2024
SAMSNeRF: Segment Anything Model (SAM) Guides Dynamic Surgical Scene Reconstruction by Neural Radiance Field (NeRF)
SAMSNeRF uses the Segment Anything Model to guide dynamic surgical scene reconstruction via NeRF. SAM-generated segmentation masks improve foreground/background separation during training, leading to sharper reconstruction of surgical tools and tissues from monocular endoscopic video.
2023
IEEE-TMI
Min-Max Similarity: A Contrastive Semi-Supervised Deep Learning Network for Surgical Tools Segmentation
Ange Lou, Kareem Tawfik, Xing Yao, Ziteng Liu, and Jack Noble
IEEE Transactions on Medical Imaging (IEEE-TMI), 2023
We propose a contrastive semi-supervised learning framework for surgical tool segmentation that leverages both labeled and unlabeled endoscopic images. The Min-Max similarity loss enforces consistency between augmented views while maximizing inter-class separation, significantly reducing annotation cost without sacrificing segmentation accuracy.
CIBM
CFPNet-M: A Light-Weight Encoder-Decoder Based Network for Multimodal Biomedical Image Real-Time Segmentation
CFPNet-M is a lightweight encoder-decoder network designed for real-time segmentation of multimodal biomedical images. By incorporating channel-wise feature pyramid modules, it achieves a strong balance between accuracy and inference speed, making it suitable for deployment on resource-constrained clinical devices.
JMI
CaraNet: Context Axial Reverse Attention Network for Segmentation of Small Medical Objects
CaraNet proposes a context axial reverse attention network for accurately segmenting small and irregularly shaped medical objects such as polyps and skin lesions. The reverse attention mechanism progressively refines object boundaries from coarse to fine, while axial attention captures long-range spatial dependencies with linear complexity.
