Advancing Drug Discovery with Multimodal AI

Selected Publications

1. Enhancing cross-dataset generalization with quasi-multimodal training and the diamond hybrid backbone

   Yuquan Xu, Taha M. Rajeh, Yutong Zhang, Li Zhang, Yuxuan Wan, Jungwoo Joo, Yuefei Wang

   Biomedical Signal Processing and Control • 2026

   • Quasi-Multimodal training
   • Medical image segmentation
   • Cross-dataset generalization
   • Dual-Path Feature Reinforcement
   • Invariant representation
   • pHybrid backbone

   open_in_new Article notes Read abstract

   Medical image segmentation faces persistent challenges arising from limited cross-dataset generalization and the prohibitive cost of large-scale expert annotation. To address these issues, we introduce a novel Quasi-Multimodal training paradigm. This data-efficient strategy integrates heterogeneous, unpaired datasets centered on a shared pathological target to mathematically reduce label space entropy, thereby compelling the model to learn scanner-invariant representations. To fully exploit this paradigm, we develop Diamond, a specialized hybrid backbone featuring a Dual-Path Feature Reinforcement mechanism. Unlike traditional single-stream architectures, Diamond employs a unique SkipRes-Connection strategy to preserve high-frequency spatial details, effectively filtering domain-specific noise while reinforcing global semantic context. Complementing this backbone, two lightweight modules are introduced: Residual Recursive Gated Convolution for high-order spatial interaction modeling and the Nested Attention System for multi-dimensional feature calibration. The core contribution is the synergistic integration of the entropy-reducing QMM strategy with the noise-resilient Diamond architecture, establishing a cohesive framework for high-fidelity segmentation under heterogeneous conditions. Extensive experiments on ten public datasets demonstrate that QMM yields a statistically significant improvement in cross-dataset performance, increasing combined IoU and Dice scores by an average of 12.54% over unimodal training. Furthermore, Diamond surpasses 14 state-of-the-art benchmarks with a 12.61% gain in cross-validation metrics while maintaining superior computational efficiency. The source code is publicly available at https://github.com/IamDerrick666/Diamond.git

2. MMTU-Net: enhancing medical image semantic segmentation with multi-level multi-scale fusion and transformer

   Xilei Wang, Yuefei Wang, Yuquan Xu, Yutong Zhang, Li Zhang

   The Visual Computer • 2025

   • Semantic segmentation
   • Vision transformer
   • U-shaped network
   • Medical image

   open_in_new Article notes Read abstract

   Semantic segmentation in medical imaging remains challenging due to issues such as semantic information loss during downsampling, excessive semantic gaps in Skip-connections, and the neglect of global information by deep networks. To address these challenges, we propose MMTU-Net, a U-shaped network augmented with a Multi-level Fusion Transformer (MFT) in the deep layers to expand the receptive field. A dual-channel attention mechanism (DCA) is introduced to reorganize channel weights, preventing insufficient edge information extraction. Furthermore, a Double-Layer Multi-Scale Feature Fusion (DMFF) module is integrated into the Skip-connections to merge images of different levels, thereby reducing feature loss and semantic gaps. Evaluated on four medical image datasets using seven metrics, MMTU-Net demonstrates superior performance compared to benchmark models. Ablation studies confirm the effectiveness of its key components.

3. A feature enhancement network based on image partitioning in a multi-branch encoder-decoder architecture

   Yuefei Wang, Yutong Zhang, Li Zhang, Yuxuan Wan, Ruixin Cao, Liangyan Zhao, Yixi Yang, Xi Yu, Zhixuan Chen, Yuquan Xu

   Knowledge-Based Systems • 2025

   • Semantic segmentation
   • Image partition
   • Multi-branch
   • Medical Image

   open_in_new Article notes Read abstract

   Semantic segmentation is of great significance in the medical field, as it can help doctors intelligently, quickly, and accurately locate key lesion areas, providing crucial support for the diagnosis, treatment, and recovery of patients. The primary reason existing segmentation networks encounter accuracy bottlenecks is that the size of lesion images limits the network’s attention to information, while also causing a lack of image information transmission between encoding and decoding. This study proposes a Multi-branch Partition Feature Enhancement Network (MPFE Net), which is essentially based on our proposed network system, the Multi-Branch Partition-Guided Decoding Network (MPD Net), using an image partition strategy. This approach divides the image at the encoding end and passes it to the decoding end, enabling parallel decoding with branches based on partitions to alleviate the problem of insufficient image information transmission. In terms of module design and optimization, under the processing idea of image partitioning, we have constructed the Bottleneck module Multi-semantic Progressive Interaction Guider (MPIG) for semantic progressive fusion. Furthermore, we have enhanced the ability to extract local information in the Vision Transformer, constructing the Multi-branched Feature Enhancement with Shared ViT (MFES ViT) to enhance control over image details. In the experiments, MPFE Net was compared with 20 models on 8 medical datasets for metric exploration, image comparison, process verification, and statistical analysis. We also discussed in detail 4 key issues. The results show that MPFE Net has better lesion universality and segmentation superiority.

4. Flattened and simplified SSCU-Net: exploring the convolution potential for medical image segmentation

   Yuefei Wang, Yuquan Xu, Xi Yu, Ronghui Feng

   The Journal of Supercomputing • 2024

   • Medical image segmentation
   • CNNs
   • Lightweight
   • Symmetric codec

   open_in_new Article notes Read abstract

   Medical image semantic segmentation is a crucial technique in medical imaging processing, providing essential diagnostic support by precisely delineating different tissue structures and pathological areas within an image. However, the pursuit of higher accuracy has led to increasingly complex architectures in existing networks, resulting in significant training overhead. In response, this study introduces a flattened, minimalist design philosophy and constructs the shallow super convolution U-shaped Net (SSCU-Net) based on this concept. Compared to the traditional four-layer U-shaped networks, SSCU-Net has a simplified two-layer structure, adhering to a lightweight research objective. On one hand, to address the issue of insufficient semantic feature extraction caused by the shallow network architecture, a parallel multi-branch feature extraction module called the super convolution block is designed to thoroughly extract diverse semantic information. On the other hand, to facilitate the transfer of critical semantic information between encoding and decoding, as well as across layers, the spatial convolution path, along with feature enhanced downsample and feature resolution upsample, are constructed. The performance of SSCU-Net was validated against 18 comparison models across seven metrics on five datasets. Results from metric analysis, image comparisons, and ablation tests collectively demonstrate that SSCU-Net achieves an average improvement of 15.9792% in the Dice coefficient compared to other models, confirming the model’s advantages in both lightweight design and accuracy. The network code is available at https://github.com/YF-W/SSCU-Net.

5. Multi-Bottleneck progressive propulsion network for medical image semantic segmentation with integrated macro-micro dual-stage feature enhancement and refinement

   Yuefei Wang, Yutong Zhang, Li Zhang, Yuquan Xu, Ronghui Feng, Haoyue Cai, Jiajing Xue, Zuwei Zhao, Xiaoyan Guo, Yuanhong Wei, Zixu Wang, Siyi Qiu, Yixi Yang, Xi Yu

   Expert Systems With Applications • 2024

   • Semantic Segmentation
   • Multi-Bottleneck
   • Semantic Progressive Propulsion
   • Feature Enhancement and Refinement
   • Medical images

   open_in_new Article notes Read abstract

   Semantic segmentation can assist doctors to locate key lesion areas intelligently, rapidly and with high precision in medical imaging, which is of great clinical significance for the diagnosis and rehabilitation of patients in various medical fields such as oncology and neurology. Existing segmentation networks face accuracy bottlenecks because enriched information at the decoding end can’t disperse over multiple layers of semantic depth. This is compounded by challenges like complex anatomical structures and varying imaging modalities. In this study, we propose a Multi-Bottleneck Progressive Semantic Segmentation Network (MBP-SSNet), which is essentially a new Baseline based on our proposal: Multi-Bottleneck based Progressive Propulsion Feature Guidance Network (MB-PPFG Net). Its Multi-Bottleneck solves the problem of coding and decoding multiplexed transmission of enriched information, and the propulsive transmission alleviates the problem of insufficient semantic fusion. In addition, under the macro–micro processing idea, MacroFocus Pre-Bottleneck Feature Enhancer (MPBFE) is proposed at the codec intersection in order to better enhance the image contextual semantic dependency association; while MicroDetail Post-Decoding Feature Refiner (MPDFR) is proposed at the tail of the network for fine-grained polishing. In the experiment, the MBP-SSNet was compared with 17 other models across 7 datasets, including diverse medical imaging scenarios, involving metric exploration, image comparison, process validation, and statistical analysis. Concurrently, 22 types of ablation analyses were conducted on it, along with detailed discussions on two critical issues. The experimental results demonstrate that MBP-SSNet achieves an average performance improvement of 0.6809 % in medical image segmentation compared to the optimal benchmark model. The official web address of MBP-SSNet is at: https://github.com/YF-W/MBP-SSNet.