Bone age assessment, a critical tool for evaluating skeletal development in children, plays a vital role in growth monitoring, disease diagnosis, and treatment. This paper proposes a highly interpretable two-stage deep learning framework for automated bone age assessment under the RUS-CHN clinical standard. Addressing the limitations of existing methods specifically, the lack of traceable clinical alignment and insufficient adaptation to Chinese children’s growth characteristics—our approach integrates anatomical priors and clinical workflows through three key innovations: (1) a RUS-CHN-compliant framework with explicit anatomical-stage modeling, (2) an enhanced YOLOv11 detector incorporating coordinate attention mechanisms for precise skeletal landmark localization, and (3) a shared-weight category-aware network that reduces model complexity while improving data utilization. Evaluated on the RSNA Pediatric Bone Age Challenge dataset, our method achieves a mean absolute error (MAE) of 7.03 months, with 80.95
With the continuous development of medicine, telemedicine has become a major trend, and the frequency of medical images being shared and interacted with online has significantly increased. In response to the poor robustness of medical image watermarking algorithms against conventional and geometric attacks, making it difficult to ensure the security of medical images, this paper proposes a zero-watermark algorithm for medical images based on FDCT-ResNet101. The watermark image is first encrypted using logical mapping chaos encryption, and then the pre-trained ResNet101 network and fast discrete curvelet transform (FDCT) transformation are used to extract the feature vectors of the medical image. The feature sequence of the medical image was obtained using a perceptual hash method, and zero-watermarking technology was applied for watermark embedding and extraction. Finally, the correlation coefficient was used to measure the correlation between the embedding and watermark extraction of the algorithm. Experimental results show that the proposed algorithm can effectively extract watermarks without altering the image and has high normalized correlation (NC) values under conventional and geometric attacks, thus demonstrating good invisibility and robustness.
Access to water is essential for human survival, animal consumption, irrigation, and various domestic and commercial uses. Unfortunately, the quality of water resources in many areas has declined and has become increasingly contaminated brought by substantial number of human activities. The evaluation of water quality entails labor-intensive and time-consuming processes like the water quality index and are often susceptible to errors. This research study aims to evaluate water quality and to identify the feature selection method which provides the highest capability in predicting water quality. Four supervised machine learning models (random forest, logistic regression. Support vector machine, and gradient boosting) were evaluated to predict water quality in both drinking and irrigation water datasets in four various model configurations assessing the effect of integrating three different feature selection methods namely: correlation method, SelectKBest method and recursive feature elimination with cross validation method. Logistic regression and support vector machine both with correlation feature selection methods appeared to be the best performing models for the drinking water dataset with the metrics ranging from 97
With the growing prevalence of digital healthcare, securing medical images has become increasingly critical. This paper introduces a robust zero-watermarking algorithm for medical images, leveraging wide residual networks (Wide ResNet) with discrete cosine transform (DCT). Initially, the original medical image was fed into the Wide ResNet model to extract feature representations, yielding a feature matrix at the output of the fully connected layer. This matrix was then subjected to DCT, followed by processing with a perceptual hash algorithm to derive a 32-bit feature vector for the image. Subsequently, the Arnold algorithm is employed to obfuscate and encrypt the watermarked image, facilitating the embedding and extraction of the watermark using zero-watermarking techniques. Experimental results demonstrate that the proposed algorithm outperforms existing methods, showing excellent resistance to both conventional and geometric attacks, as well as strong robustness against composite attacks.
Sleep disorders have become a global health problem, and long-term sleep disorders pose a serious threat to an individual’s physical health, such as increasing the risk of cardiovascular disease, metabolic disorders and neurodegenerative diseases. So its early recognition and intervention is very important. At present, the diagnosis of such diseases is mainly made by the direct clinical assessment of the patient by the surgeon. The diagnosis requires a lot of time and energy, and the diagnosis process often has a certain lag, and the patient often needs to show certain symptoms to be diagnosed. In addition, the subjective evaluation of doctors is also an important problem in determining the diagnosis result, which lacks objectivity and automatic repeatability. In order to solve the above problems, the objective of this paper is to develop an intelligent risk prediction system based on image processing. In this paper, a multi-modal fusion method using tongue manual features and deep attention features is proposed, which can mine more multi-dimensional depth features to assist risk prediction while ensuring interpretability. The experimental results show that the multi-modal feature fusion method is superior to the traditional single processing method.
Electroencephalogram (EEG) signals have emerged as a promising tool for the diagnosis of major depressive disorder (MDD). However, significant individual differences in EEG data limit the performance of cross-subject detection models. To address this challenge, we propose a novel degree interaction domain adaptation (DIDA) model designed to enhance the accuracy of MDD detection in cross-subject EEG scenarios. First, this study proposes a new cross-subject data interaction domain adaptation method featuring degree-weighted functionality to extract shared features from EEG signals across different subjects. Additionally, a degree-based weighting mechanism is proposed in the data interaction module, effectively reflecting the importance of different nodes in the latent space. Finally, a sinusoidal function-based dynamic weight adjustment strategy is proposed, enabling the model to adapt data interaction intensities dynamically during different training stages, optimizing the balance between interaction and detection. Experimental results demonstrate that DIDA achieves robust MDD detection performance.
The process of drug discovery involves searching large databases and analyzing millions of (secret) documents. More specifically, drug development projects result in complex folder structures. Existing systems lack complex inquiry abilities. Therefore, the search process is manual, time-consuming, and inefficient. Recent advancements in machine learning and AI provide opportunities to augment the search process. The aim of this paper is to report the results and lessons learned from a pilot project at a global pharmaceutical manufacturing company involving the development of a prototype for augmenting search activities. The prototype was developed using agile software development and tested in three iterations. The AIDAF for design thinking approach and intelligence amplification design canvas were used to analyze the developed prototype. Focus group meetings were held to evaluate the results, capture user experiences, and lessons learned. The pilot resulted in an operational front-end application and integration with an internal database. Text from 800K of 1.5M documents and corresponding folder structures was extracted to train machine learning models that can predict folder locations for new documents. Although the initial accuracy improved from 30
Artificial intelligence (AI) has been recognized by the World Health Organization for its transformative potential in addressing global reproductive healthcare challenges, including inequitable access to monitoring and treatment and limited diagnostic precision. AI offers significant promise in enhancing diagnostic accuracy enabling data-driven decision-making for personalized preventive and therapeutic interventions. However, its deployment also raises ethical and operational concerns, such as data privacy risks, algorithmic bias, legal complexities, cultural sensitivity, overreliance on AI-generated recommendations, and the potential deskilling of clinicians. Addressing these challenges requires inclusive frameworks for responsible integration. Moreover, AI-driven digital transformation must align with the broader call for sustainable innovation outlined in the United Nations’ 2030 Agenda and European Union regulations defining the requirement for safe and secure integration standards. This research explores pathways for responsible and sustainable AI adoption in reproductive healthcare while mitigating associated risks. It introduces the (H)iCARE framework, which advocates for (1) human-centric hybrid models that integrate AI-driven innovations with clinical expertise to ensure balanced innovations. The framework also (2) embraces a broader, humanity-oriented perspective to ensure inclusivity beyond a limited subset of stakeholders, and (3) fosters a learning-driven approach that prioritizes continuous skill development to prevent cognitive complacency. While developed in the context of reproductive healthcare, its principles extend across the healthcare sector, providing a foundation for AI-integrated information system design and a roadmap for ethical, sustainable advancements, additionally fostering discussion on future research priorities.
Longitudinal analysis is crucial in different areas of medicine. Finding similarities between patients’ data over time (longitudinal data) allows the definition of subgroups of similar patients, unveiling new aspects of disease progression and fostering new therapies specifically tailored for each group. Longitudinal clustering has become an interesting tool for grouping subjects and analyzing data trajectories over time. These trajectories are often represented as time series. In the healthcare domain, the analysis of longitudinal data poses some challenges due to their complex representation and inherent nature, such as high dimensionality and variability. To address these issues, we propose a Long-Short-Term Memory (LSTM) based autoencoder to enhance the k-means longitudinal clustering by encoding high-dimensional and high-variable longitudinal data into a compressed representation. The proposed approach was evaluated on synthetic datasets, showing an improvement in performance compared to using only longitudinal k-means clustering. Moreover, it was applied to a real longitudinal dataset of Alzheimer’s Disease (AD) patients to also assess the efficacy of the proposed approach in real longitudinal health data.
Emerging evidence suggests a potential link between Long COVID and neurological disorders, including Parkinson’s Disease (PD). This study explores brain-network alterations in Long COVID and PD patients through a novel telecommunication-based brain model. Borrowing language and methods from telecommunications, we build a novel theoretical channel model for the brain network. By examining functional and structural brain connectivity through the proposed channel model, we aim to identify overlapping and distinct patterns of brain network disruption. Hence, considering real data, we compute the network alterations occurring in patients affected by Long COVID and PD, comparing them with healthy controls. The comparison reveals common alterations in regions associated with dopamine alterations, and differences for memory and shape recognition. Our findings highlight potential shared pathways between Long COVID and PD, contributing to a deeper understanding of how viral infections may influence neurodegenerative processes and offering insights for targeted interventions.
In today’s fast-paced medical industry, the effective management of documents is crucial for ensuring patient safety, regulatory compliance, and operational efficiency. Combining the provision of healthcare services in the clinic and at patient’s home requires the creation of a unified data exchange system between patients, family doctors and doctors in hospitals. Managing a medical organization involves navigating complex regulations and maintaining compliance with ever-changing healthcare laws. This article explores an architectural approach to the implementation of Electronic Document Management Systems (EDMS) specifically tailored for medical companies. A case study of a medical company offering house call services reveals the features encountered during the integration of EDMS. Ultimately, our findings underscore the significance of adopting a structured architectural framework that enhances document management, promotes collaboration within a matrix hierarchy, and leads to improved patient care in the unstable domain of home healthcare services.
Medical images are crucial for healthcare diagnosis, treatment planning, and research. However, during transmission or storage, the integrity and authenticity of these images may be compromised, potentially leading to misdiagnosis or privacy breaches. This paper proposed a novel zero-watermarking algorithm, RegNet-DCT, which leveraged the state-of-the-art convolutional neural network architecture RegNet, combined with Discrete Cosine Transform (DCT). By utilizing RegNet, both local and global features of medical images were effectively captured. These features were then transformed into the DCT domain, where the watermark information was embedded via a perceptual hash and logistic map. The RegNet-DCT-based zero-watermarking algorithm ensured that the watermark was invisible to the human eye, and the original image is not required for high quality watermark extraction. Experimental results on various types of medical images show that the proposed watermarking algorithm exhibited superior robustness against both conventional and geometric attacks.
Thanks to the rapid growth in artificial intelligence, speech recognition is a task that allows for increasingly precise transcription of spoken language into text. Unfortunately, one sector that has been overlooked in this development is disordered speech, affecting individuals with dysarthria, apraxia, aphasia, or other conditions impacting clear speech. For such individuals, a solution capable of recognizing and transcribing their speech into text would be especially beneficial, as it would enable more effective communication in their daily lives. The main challenge in creating such an AI (Artificial Intelligence) model is the lack of training data in audio recordings from individuals with disordered speech. This work aims to fill in this research gap by creating a data-collection application along with a dedicated database and initiating work on an AI model specializing in recognizing disordered speech in Polish. Therefore, the first step is to develop a web application that will be available to speech therapists and rehabilitation centers to start collecting audio recordings. Through strategic collaboration with institutions and professionals specializing in treating and rehabilitating disordered speech, a Polish-language database is created to train the speech recognition model.
Background: Breast cancer (BC) can be classified into four PAM molecular subtypes. Accurate subtype classification is crucial to predict prognosis and support treatment decisions but has multiple limitations. This study explores transfer learning (TL) to improve subtype classification in data-limited scenarios. Objective: Assess how TL improves subtype predictions using gene expression. Methods: We developed a three-stage framework where we trained (1) a foundational model to distinguish cancerous from healthy samples, (2) a subtype classifier to predict subtypes and (3) a fine-tuned subtype classifier leveraging the pretrained weights from the foundational model. Results: TL substantially improved subtype classification, achieving a 0.777 final accuracy, in comparison to a 0.562 accuracy achieved using randomly initialized weights. Conclusion: Our results demonstrate that TL enhances molecular subtype prediction, and it is an approach that deserves to be explored in further depth in precision oncology settings.
This work introduces L-CNN, a convolutional neural network designed for the detection of leukemia cells, enhancing pediatric cancer diagnostics by distinguishing cancerous from healthy cells. The model was trained on approximately 15,000 images from 118 subjects, sourced from a Kaggle challenge dataset. Leveraging EfficientNetV2B3, a scaled-down variant of EfficientNetV2-M, L-CNN achieves a test accuracy of 96.87
Biofeedback refers to training methods that assist individuals in controlling specific involuntary bodily functions, such as heart rate, breathing, and brain wave activity, among others. It is becoming increasingly popular as a complementary approach to therapies for children with autism (ASD) because it trains children in self-regulation skills, helps reduce anxiety, and promotes relaxation. Many individuals with autism become anxious and frustrated when asked to participate in social-communicative activities. Sometimes, children exhibit external behaviors that show their frustration. However, it is not always evident that they might be distressed because of their difficulties in language and social tasks. This article illustrates the use of biofeedback in therapeutic interventions for children with Autism Spectrum Disorder through a case study. Observing a child’s physiological changes in real-time during speech therapy enhances understanding of tasks or demands that may elevate the child’s stress levels. This, in turn, provides therapists with a clearer view of the child’s communicative competence and insight into how to modify therapeutic strategies.
Pulmonary diseases, including lung and respiratory diseases, are a significant global health issue, causing nearly 4 million deaths in 2019. Chronic obstructive pulmonary disease is the third leading cause of death worldwide, with lung cancer being the leading cause. In 2021, the EU’s respiratory disorders death rate was 65.5 deaths per 100,000 inhabitants. This study applies Deep Learning models to classify RX thoracic images of various lung diseases, achieving high accuracy values (97/98
Carotid Atherosclerosis (CAS), a significant risk factor for cardiovascular disease and stroke, necessitates accurate early detection. This study evaluates machine learning models for classifying CAS severity based on Doppler ultrasound features. Using MATLAB, we processed ultrasound signals to extract critical biomarkers including spectral entropy, peak frequency shifts, spectral energy, and blood flow velocity estimations. We compared three classification algorithms—Support Vector Machine (SVM), Random Forest (RF), and K-Nearest Neighbors (KNN)—optimized through hyperparameter tuning with five-fold cross-validation. To enhance model generalization, we implemented feature standardization and noise augmentation techniques. Results demonstrated that SVM and Random Forest models achieved equivalent superior performance (83.33
Medical emergencies, such as suddenly declining vital parameters during surgical procedures, can induce acute stress in attending staff and may contribute to errors in emergency management, posing risks to patient safety. Anesthesiologists, who frequently encounter high-stress clinical scenarios, are particularly susceptible to health issues, including burnout. To mitigate diagnostic and treatment errors during surgical emergencies, an algorithm-based cognitive aid, eGENA, has been developed by incorporating human factors principles from aviation. Initially released in 2020, eGENA has since been introduced in several pilot hospitals across Germany. Preliminary findings have demonstrated positive user acceptance and improvements in various performance indicators during surgical procedures. Due to the inhomogeneous nature of real emergencies in the operating room (OR) setting, data acquisition is difficult. This study focuses on the use of a within-subjects experimental design through standardized in-situ simulation scenarios in the OR to collect subjective stress measures from 20 physicians and nurses with and without eGENA support. The data obtained suggests that the use of simulation-based emergency cases can aid to investigate the impact of eGENA on individual stress levels of various teams during surgical procedures.
Epilepsy is a neurological disorder affecting millions worldwide, often requiring accurate and timely detection for effective treatment. Recent advancements in deep learning have significantly improved automated classification, using for instance time-series as input data. In this paper, we propose an approach exploiting three different long short-term memory networks for possible epileptic seizure classification. The proposed method is able to classify a electroencephalogram data into one of the following categories: eyes open state, eyes closed state, healthy brain regions, tumor-affected brain regions and seizure activity. Experimental results highlight the effectiveness of the proposed approach in distinguishing epileptic (related to the seizure activity category) and non-epileptic (related to the eyes open state, eyes closed state, healthy brain regions, tumor-affected brain regions categories) electroencephalogram data, thus showing the effectiveness of the proposed method.
