In this paper we propose a low-cost 3D segmentation framework which can realize HR segmentation with LR input. Thus, patch-sampling is no longer needed and the model can have the global context when training. Experiments show that our model outperforms patch-based methods with a 4x higher speed.

In this paper we propose a new Vision Transformer module to simutaneously model local and global dependecies. The self-attention is split into fine-grained local SA and coarse-grained global SA with a gaussian mask further added on it. Experiments show that our model achieved sota performance on two datasets.

In this paper we propose a simple yet efficient pure-MLP multimodal fusion technique named CubeMLP. CubeMLP fuses the multimodal input in sequence, modality and channel dimensions repectively, achieving SOTA performance on CMU-MOSI, CMU-MOSEI and AVEC2019 with a much lower complexity.

This paper is the extended version of our MICCAI 2021 paper. The main improvements include a selective cropping algorithm for HR guidance patch, the use of multi-scale encoders, and test-time refinement based on the existing LR-HR image pair. This paper is finished in January 2022.

In this paper, we propose to solve the information propagation gap between the embedder and classifier in MIL with an iteratively coupled method. This elegant method view the entire MIL pipeline as an EM optimization problem, and is highly generalizable. We adopt ICMIL on different MIL backbones and achieved consistent improvements. The main branch of the repository holds the code for this method.

In this extended journal version of ICMIL, we propose to use the instance aggregator in MIL backbones to generate confidence score for each instance's pseudo label, thereby improving the effectiveness of the embedder fine-tuning. A bag-level augmentation method is also added for the MIL training to realize a even higher performance. Check our code in the alternative branch of the repository.

In this work, we propose a novel referring segmentation network named SLViT. SLViT introduces an effective encoder block for simutaneous feature extraction and cross-modal attention. Then, it proposes LMFA to exploit the multi-scale information to generate more comprehensive outputs. Furthermore, a novel URCE module is also proposed to find and refine the high uncertainty tokens based on the cross-scale information. SLViT achieves SOTA performance on three datasets.

MCKD combines UDA and DG with Federated Learning to achieve a higher performance while preserving the users' privacy. collaborative distillation strategy using multiple local models and a centralized model. Experimental results demonstrate its competitive performance compared to state-of-the-art methods on four datasets.

In this work, we propose a multi-scale vision-language model named HSVLT for multi-label image classification. HSVLT features IVLA and CSA, of which the former is for joint updating of visual features, linguistic features, and multi-modal features considering interactive cross-modal cues, and the later is for leveraging complementary information from different scales for decision-making purposes. HSVLT achieves sota performance on three different datasets.

In this paper, we propose a novel patch-free 3D medical image segmentation network named VolumeFormer. It features Adaptive Decomposition, which is the inproved learnable version of Holistic Decomposition, and Shared-Weight Volumetric Transformer Block, a lightweight 3D transformer block that can realize cross-scale weight-sharing. The experimental results demonstrate its efficiency and effectiveness.

In this paper, we propose a many-to-one regression transformer, M2ORT, for Spatial Transcriptomics prediction. We also propose ITMM and ICMM to decouple the multi-scale modeling process in M2ORT, leading to notable savings in computational cost without compromising performance.