Ocular myasthenia gravis (OMG) is a challenging condition to diagnose, with Cogan's lid twitch (CLT) serving as a key clinical sign. Early and accurate diagnosis of OMG is crucial for timely intervention and improved patient outcomes. However, current diagnostic methods for CLT rely primarily on visual assessment by clinicians, which is inherently subjective and prone to interobserver variability, highlighting the need for more objective diagnostic tools. This pilot study was the first to investigate the potential of video-based eyelid tracking for simplified diagnosis of Cogan's lid twitch (CLT), a key sign of ocular myasthenia gravis (OMG). The importance of this research lies in its potential to enhance diagnostic accuracy, enable earlier treatment initiation, and improve the overall management of OMG patients. We analyzed pixel value changes and eyelid position in a video recording of CLT, employing preprocessing techniques to stabilize footage and reduce lighting artifacts. Eyelid tracking was performed using contour detection and polynomial fitting. Our results showed detectable variations in eyelid position during CLT occurrences, though with inconsistent accuracy. Challenges included distinguishing subtle CLT movements from normal blinks and addressing issues like double eyelids. While further refinement is needed, this research suggests the potential of video-based tracking as a non-invasive tool for OMG diagnosis, offering significant clinical implications.Clinical Relevance- This study suggests the potential for simplified diagnosis of Cogan's lid twitch through video-based tracking, indicating significant clinical implications.

Laryngo-pharyngeal cancer(LPC) is a high-mortality malignancy whose early diagnosis is critical for effective treatment. However, traditional diagnostic methods rely heavily on laryngologists' expertise which will cause possible missed diagnoses or repeated biopsies. To address this, we propose the Laryngo-Pharyngeal Cancer Diagnosing System (LPC-SYS), an AI-powered system that leverages YOLO-based object detection models for real-time LPC detection from endoscopic images. LPC-SYS uses a microservices architecture to ensure scalability and efficient task management, enabling seamless integration into clinical workflows. Meanwhile, extensive experiments conducted on two real private LPC datasets demonstrate that LPC-SYS, utilizing YOLO models, achieves competitive results. With its high diagnostic accuracy and minimal delays, LPC-SYS is set to be deployed in the clinical settings of the First Affiliated Hospital, Sun Yat-sen University. This implementation will provide a reliable and scalable solution for LPC diagnosis.

The treatment of bulky tumors with conventional radiotherapy is limited by the need for higher therapeutic doses, which can increase toxicity to healthy tissues and fail to exploit tumor heterogeneity. In this context, Lattice radiotherapy emerges as an innovative approach, focusing on heterogeneous partial irradiation that alternates areas of high dose, known as vertices, with lower dose zones. This method promotes immunogenic cell death in the vertices, releasing antigens and inflammatory cytokines that enhance immune system activation while reducing radiation exposure to at-risk organs and modulating the dose in areas of the tumor that are less radio-sensitive, such as hypoxic or necrotic zones. To facilitate the implementation of this technique, a Lattice-based algorithm has been developed to automate the generation and three-dimensional distribution of the vertices within the tumor volume, using DICOM RT files that contain CT images and anatomical segmentations. The developed algorithm allows for the adjustment of parameters such as the diameter and spacing of the vertices, as well as the ability to remove or modify their locations, thus optimizing the protection of surrounding organs. All of this is presented in an intuitive and interoperable graphical user interface, enabling the integration of the generated spheres into radiotherapy planning systems.Clinical RelevanceThis Lattice-based algorithm offers radiation oncologists a clinically significant tool for optimizing treatment by tailoring high-dose sphere distributions within the target tumor volume. It determines sphere diameters, defines boundaries between spheres and the target contour, and allows for adjustments such as redistributing, resizing, or removing spheres. Notably, the algorithm facilitates these adjustments not only in two-dimensional planes, where such modifications are relatively straightforward, but also in three-dimensional space, addressing the complexities that arise when working across multiple axes. This enhanced precision enables more effective tumor coverage while sparing healthy tissue, leading to improved therapeutic outcomes and enhanced patient quality of life.

Cancer therapy-related cardiac dysfunction (CTRCD) is an increasingly significant concern due to cardiac function deterioration caused by anticancer drug side effects. While echocardiography is the conventional diagnostic method for CTRCD, its accuracy heavily depends on operator expertise and the procedure is both time-consuming and costly. Electrocardiogram (ECG), being more accessible and easier to measure, presents a promising lower-cost alternative. In this paper, we propose a deep learning model for CTRCD detection from ECG signals. Our model integrates convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to capture both local and global ECG features, while incorporating an attention mechanism to comprehensively learn feature importance. To enhance model interpretability, we visualize the attention weights to identify ECG features that significantly contribute to the classification decision. Through extensive ablation studies using standard 12-lead ECG data, we demonstrate the effectiveness of our proposed architecture. This work is expected to contribute to the development of cost-effective and reliable diagnostic tools for monitoring cardiac side effects during cancer treatment.

Prostate cancer (PCa) diagnosis often relies on biopsies, which can lead to unnecessary procedures and complications. Federated learning (FL) offers a privacy-preserving approach for training predictive models across hospitals without sharing sensitive patient data. In this study, we evaluate the feasibility of FL for PCa risk prediction by benchmarking different training strategies, including local, federated models, as well as free-riding (FR) on federated models. Using real-world heterogeneous datasets from 19 hospitals, we analyze the impact of data diversity and consortium size on predictive performance. Our results show that while FL improves model generalizability, local models often perform comparably, making direct participation in FL less beneficial for large hospitals. However, a small consortium of high-data-quality institutions could collaboratively develop robust models for broader clinical use. We discuss the practical implications of FL in healthcare and propose strategies for sustainable deployment in real-world hospital networks.

Lung cancer remains the most prevalent cancer worldwide, yet challenges like late-stage diagnosis and limited treatment options continue to pose serious public health concerns. The persistently low five-year survival rate highlights the urgent need for more effective diagnostic and prognostic methods. In response, we propose MMLCA, a multimodal learning framework designed to predict EGFR mutation types and survival outcomes in lung cancer patients. MMLCA integrates diverse medical data sources, including lung CT images, clinical notes, laboratory results, and basic information, using a hierarchical cross-attention mechanism that captures complementary insights across modalities. By leveraging the full spectrum of available patient data, MMLCA significantly enhances predictive accuracy. Our experiments show that MMLCA consistently outperforms traditional approaches, suggesting it could help clinicians make more accurate predictions and support more personalized treatment decisions.

Breakthroughs in Artificial Intelligence (AI) are reshaping oncology, facilitating early cancer detection, more accurate diagnosis, and personalized treatment. However, its clinical adoption is hindered by black-box models' opacity, raising concerns about reliability, accountability, and ethical implications. Although explainable AI (XAI) helps mitigate these issues, it often faces challenges, such as low fidelity, fairwashing, and vulnerability to spurious correlations and biases. Causal eXplainable AI (CXAI) aims to address these shortcomings by leveraging causal inference, thereby improving model robustness, fairness, and clinical relevance. While recent studies have explored causality, explainability, and AI in healthcare, they largely remain conceptual and lack a comprehensive synthesis in oncology. This paper fills this gap by providing the first comprehensive review of CXAI for cancer diagnosis. We summarize current applications across multiple cancer types, identify pressing challenges, and propose future directions. Our findings highlight CXAI's potential to make AI-driven oncology more trustworthy, transparent, and effective.Clinical relevance-Causal Explainable AI (CXAI) significantly enhances clinical decision-making by capturing causal relations in medical data, thus improving predictive accuracy and aligning AI insights with clinical reasoning. By increasing transparency and trust, CXAI facilitates AI adoption in oncology, enabling informed diagnostic decisions and personalized patient care.

Endoscopy plays a crucial role in the early diagnosis of colon cancer. The manual processing of images by skilled endoscopists is time-consuming, making automatic image classification highly valuable. We propose a novel multi-label classification method that integrates complementary learning from both local and global approaches. The model comprises a Swin Transformer branch for global feature extraction and a modified VGG16-based CNN branch for local feature analysis. The learning capability of the CNN branch is enhanced by concatenating a saliency map and the prediction of a texture feature vector through a multi-task learning framework. The proposed method outperformed state-of-the-art techniques, achieving an F1-score of 96.08% and an accuracy of 96.06% on the classification of the Kvasir-v2 dataset of endoscopic images.Clinical Relevance-Experimental results demonstrate the superiority of the proposed model for classifying endoscopic images, paving the way for enhanced diagnostic performance in clinical settings.

Breast cancer is a significant health risk for middle-aged and elderly women, with high incidence and mortality rates. Accurate diagnosis and treatment rely on both mammography and breast MRI images, which provide complementary diagnostic information. Combined observation of these images helps localize lesions and plan treatments. However, direct mapping of information between 2D mammography and 3D MRI images is challenging. We propose a 2D/3D registration framework for breast MRI-Mammography images-BreastBioMorph (BBMorph). A biomechanical finite element model based on 3D MRI reconstructions and hyperelastic Mooney-Rivlin material simulates breast tissue and mammography compression, providing coarse alignment between deformed MRI volumes and mammography. An Edge-Guided unsupervised registration network integrates breast edge constraints into a pyramid structure, refining deformation fields progressively. The effectiveness of the registration framework is validated with clinical breast data. Experimental results are compared with those of state-of-the-art cross-modal registration approaches. The proposed method achieves promising results, with a Dice coefficient of 89.05%, MI of 0.497, SSIM of 0.815, ASSD of 17.11, and %|J_φ|≤0 of 0.104. The method provides a valuable reference for the registration of cross-modal breast images.

Non-muscle-invasive bladder cancer (NMIBC) is a relentless challenge in oncology, with recurrence rates soaring as high as 70-80%. Each recurrence triggers a cascade of invasive procedures, lifelong surveillance, and escalating healthcare costs-affecting 460,000 individuals worldwide. However, existing clinical prediction tools remain fundamentally flawed, often overestimating recurrence risk and failing to provide personalized insights for patient management. In this work, we propose an interpretable deep learning framework that integrates vector embeddings and attention mechanisms to improve NMIBC recurrence prediction performance. We incorporate vector embeddings for categorical variables such as smoking status and intravesical treatments, allowing the model to capture complex relationships between patient attributes and recurrence risk. These embeddings provide a richer representation of the data, enabling improved feature interactions and enhancing prediction performance. Our approach not only enhances performance but also provides clinicians with patient-specific insights by highlighting the most influential features contributing to recurrence risk for each patient. Our model achieves accuracy of 70% with tabular data, outperforming conventional statistical methods while providing clinician-friendly patient-level explanations through feature attention. Unlike previous studies, our approach identifies new important factors influencing recurrence, such as surgical duration and hospital stay, which had not been considered in existing NMIBC prediction models.