The integration of artificial intelligence (AI) into medical imaging has revolutionized healthcare by enhancing diagnostic accuracy, optimizing treatment planning, and improving patient outcomes. AI-driven techniques in medical image processing, classification, segmentation, and treatment planning have enabled the automated analysis of complex medical data, facilitating early disease detection and personalized medicine. Despite these advancements, challenges such as data heterogeneity, interpretability, and integration into clinical workflows persist. This special issue aims to gather cutting-edge research and developments in AI applications for medical imaging, fostering collaboration among researchers, clinicians, and industry professionals to address these challenges and propel the field forward. This issue collects four excellent articles that have been reviewed and highly recommended by the editors and reviewers.
The first paper, titled “Individualized Treatment Effect Inference of HNC with Multimodal Data” and authored by Yawen Wei, Zhen Li, Jonghye Woo, Jinsong Ouyang, Georges El Fakhri, and Xiaofeng Liu, addresses the critical challenge of estimating individualized treatment effects (ITE) for personalized medicine. The authors develop an end-to-end deep causal learning framework that incorporates multimodal patient data (CT images and clinical variables) for accurate ITE inference in a retrospective head and neck cancer (HNC) study. To effectively fuse heterogeneous data and mitigate treatment selection bias, the paper proposes a bi-stage adaptive instance normalization (Bi-AdaIN) mechanism and a mutual information-based disentanglement strategy. Evaluated on the large-scale RADCURE data set, the proposed method demonstrates a significant reduction in bias-adjusted treatment effects compared to conventional methods, offering a promising approach for counterfactual reasoning in clinical oncology.
The second paper, titled “Radiograph Super-resolution with Pixel-level Masked D-MNet Transformer,” authored by Yongsong Huang, Tomo Miyazaki, Zhengmi Tang, Kaiyuan Jiang, Yaohou Fan, Dongming Yu, and Shinichiro Omachi, focuses on enhancing the quality of medical X-ray images. The authors propose a novel architecture, the Depth-wise convolution Multiscale Network Transformer (D-MNet Transformer), specifically designed to represent and extract fine-grained local structures such as skeletal features in radiographs. Furthermore, they introduce a task-specific pixel-level masking strategy to improve model robustness when training on small sample data sets. Experimental results indicate that this method achieves state-of-the-art performance in peak signal-to-noise ratio, recovering critical details essential for accurate diagnosis.
The third paper, titled “LG-NuSegHop: A Local-to-global Self-supervised Pipeline for Nuclei Instance Segmentation” and authored by Vasileios Magoulianitis, Catherine A. Alexander, Jiaxin Yang, and C.-C. Jay Kuo, presents a fully unsupervised pipeline for nuclei segmentation in histology images. To overcome the reliance on expensive manual annotations and the “black-box” nature of traditional deep learning models, the authors introduce a transparent, data-driven feature extraction model named NuSegHop, based on the Green Learning paradigm. The pipeline integrates local processing operations with global post-processing to generate high-quality segmentations without using manually annotated training data. The method demonstrates strong generalization capabilities across multi-organ data sets, outperforming several self-supervised and weakly supervised methods while remaining competitive with fully supervised approaches.
The fourth paper, titled “DCTNet: Densely Contextual Transformer Networks for Diagnosis of Cervical Lesions,” authored by Ping Li, Tianxiang Xu, Yao Liu, Yuling Fan, YaTing Hong, Jing-Ming Guo, and YuChun Lv, proposes a novel network for the hierarchical assisted diagnosis of cervical lesions from colposcopy images. The authors design a Densely Contextual Transformer Network (DCTNet) that benefits from a specialized DCT module to combine the global contextual capabilities of Transformers with the local contextual strengths of CNNs. Additionally, a Multi-spectral Channel Attention module is introduced to enrich feature representation by utilizing multiple frequency components. The study demonstrates that DCTNet achieves diagnostic performance comparable to expert physicians, highlighting its potential to alleviate workload in resource-constrained clinical settings.
The papers featured in this special issue encompass a broad spectrum of topics within the field of AI for medical image analysis, ranging from causal inference and super-resolution to segmentation and computer-aided diagnosis. They not only shed light on the technical innovations required to handle complex medical data but also present practical solutions for clinical applications. We anticipate that this special issue will encourage researchers to explore new directions and inspire further advancements in AI-driven healthcare. Finally, we extend our heartfelt gratitude to all the reviewers for their dedicated collaboration and valuable feedback.
