SafeClick: Error-Tolerant Interactive Segmentation of Any Medical Volumes via Hierarchical Expert Consensus

Yifan Gao#, Jiaxi Sheng#, Wenbin Wu, Haoyue Li, Yaoxian Dong, Chaoyang Ge, Feng Yuan, Xin Gao

MICCAI 2025

Project

• SafeClick is an error-tolerant interactive segmentation approach that uses hierarchical expert consensus to handle imperfect user prompts. As a plug-and-play module compatible with foundation models like SAM 2, it significantly improves segmentation accuracy and robustness across 15 public datasets.

A Composite Alignment-Aware Framework for Myocardial Lesion Segmentation in Multi-sequence CMR Images

Yifan Gao, Shaohao Rui, Haoyang Su, Jinyi Xiang, Lianming Wu, Xiaosong Wang

MICCAI 2025

Project

• CAA-Seg addresses spatial misalignment and intensity variations in multi-sequence CMR images through selective slice alignment and hierarchical feature alignment. It achieves superior performance in myocardial lesion segmentation, particularly for myocardial infarction, on a large-scale dataset.

Prior-driven refinement network for small organ-at-risk segmentation in head and neck cancer

Taibao Wang#, Yifan Gao#, Bingyu Liang, Qin Wang

Engineering Applications of Artificial Intelligence

Project

• PRNet is a prior-driven refinement network that leverages foundation models and domain-specific knowledge for small organ-at-risk segmentation. It incorporates a prior encoder and mask refinement transformer to improve accuracy, demonstrating superior performance on multiple public datasets.

Dino U-Net: Exploiting High-Fidelity Dense Features from Foundation Models for Medical Image Segmentation

Yifan Gao, Haoyue Li, Feng Yuan, Xiaosong Wang, Xin Gao

arXiv 2025

Project

• Dino U-Net exploits high-fidelity dense features from the DINOv3 foundation model for medical image segmentation. It features a fidelity-aware projection module to preserve feature quality, achieving state-of-the-art performance and scalability across diverse medical imaging datasets.

Multimodal Large Language Models Address Clinical Queries in Laryngeal Cancer Surgery: A Comparative Evaluation of Image Interpretation Across Different Models

Bingyu Liang#, Yifan Gao#, Taibao Wang, Lei Zhang, Qin Wang

International Journal of Surgery

Project

• Six multimodal large language models (MLLMs) were evaluated across 6 image types, 169 images, and 1084 open-ended clinical questions in laryngeal cancer surgery. Advanced MLLMs demonstrate high accuracy (up to 79.43%) in interpreting diverse image modalities, with commercial models outperforming open-source alternatives. MLLMs show potential to enhance clinical decision-making across the surgical timeline of laryngeal cancer, from preoperative planning to post-operative care.

DeSAM: Decoupled Segment Anything Model for Generalizable Medical Image Segmentation

Yifan Gao, Wei Xia, Dingdu Hu, Wenkui Wang, Xin Gao

MICCAI 2024

Project

• DeSAM addresses the performance degradation of SAM in automatic segmentation scenarios. By decoupling the mask generation process from prompts and introducing prompt-relevant IoU and prompt-decoupled mask modules, DeSAM achieves state-of-the-art performance in two public datasets.

MBA-Net: SAM-driven Bidirectional Aggregation Network for Ovarian Tumor Segmentation

Yifan Gao, Wei Xia, Wenkui Wang, Xin Gao

MICCAI 2024

Project

• We developed MBA-Net, a novel architecture for ovarian tumor segmentation that integrates the Segment Anything Model (SAM) with domain-specific knowledge through bidirectional feature aggregation. MBA-Net demonstrates superior segmentation accuracy, robustness to tumor heterogeneity, and strong generalization capability across different imaging modalities and clinical sites.

An Anatomy-aware Framework for Automatic Segmentation of Parotid Tumor from Multimodal MRI

Yifan Gao#, Yin Dai#, Fayu Liu, Weibing Chen, Lifu Shi

Computers in Biology and Medicine

Project

• We propose PT-Net, a Transformer-based multimodal fusion network that effectively integrates information from three MRI modalities. By incorporating a novel anatomy-aware loss function, PT-Net significantly improves segmentation accuracy by distinguishing parotid tumors from similar surrounding anatomical structures, outperforming existing state-of-the-art methods.

The Professional Go Annotation Dataset

Yifan Gao, Danni Zhang, Haoyue Li

IEEE Transactions on Games

Project

• We present PAGE, the first large-scale dataset of professional Go games with extensive annotations. Spanning over 70 years, it contains 98,525 games played by 2,007 professional players. The dataset includes comprehensive game-level metadata and detailed in-game statistics generated by the KataGo engine, enabling advanced research in game analysis and AI.

Efficiently Mastering the Game of NoGo with Deep Reinforcement Learning Supported by Domain Knowledge

Yifan Gao#, Lezhou Wu#

Electronics

Project

• We propose NoGoZero+, an enhanced AlphaZero-based agent for the game of NoGo. By introducing innovative training strategies, it achieves a 6x speedup in training and defeats the original AlphaZero agent with significantly less data (20k vs 120k games). The program was the runner-up in the 2020 China Computer Game Championship.