Keywords: Artificial Intelligence, Deep Learning, Health Monitoring
Abstract: Healthcare systems generally constitute a combination of healthcare facilities and providers that collaborate in order to offer a complete range of healthcare services. As the primary providers of healthcare, hospitals hold an essential place in the healthcare system. This study introduces three novel automated Artificial Intelligence (AI) systems by using the robotic platform of the TIAGo robot: (i) an image-based system that monitors the food consumption of the patients and provides information regarding their energy intake to prevent malnutrition; (ii) a rehabilitation system that monitors patients as they perform prescribed exercises to prevent loss of mobility, and (iii) a monitoring system for the correctness of the oxygen therapy in terms of the proper positioning of the oxygen mask to prevent hypoxia complications. A well-designed observational clinical study is planned to evaluate and validate the utility and effectiveness of the proposed AI systems in improving patient stay in the hospital. For the abovementioned systems, three novel datasets were created.

Keywords: Data Science, Digital Health, Healthcare Informatics
Abstract: Depression is a disorder afflicting individuals in terms of low mood and loss of pleasure or interest in normal activities, thereby affecting their quality of life. Experience sampling method (ESM) presents a tool to investigate the interplay between depression, personality traits and affective emotional activation in light of their individual behavior and response to daily stimuli. Multivariable logistic regression was used to explore the relationship between self-reported i) characteristics prior to ESM: with ii) emotions during ESM with depression. Clinical Relevance—A total of 142 participants with moderate and severe depression were observed. Our findings suggest low self-esteem before ESM (odds ratio [OR], 0.787; 95% confidence interval [CI] 0.693-0.895; P = 0.001) and negative affect during ESM, adjusted for prior covariates (OR, 1.077; 95% CI 1.052 to 1.103; P = 0.003) are more pronounced in severely depressed individuals.

Keywords: Health Data, Diagnostics, Machine Learning
Abstract: Chin EMG is traditionally used in sleep studies to determine rapid eye movement (REM) sleep. During REM chin EMG drops to its lowest levels. Some recent works have showed good accuracy in detecting obstructive sleep apnea (OSA) utilizing the chin EMG. In this work we studied the behavior of the chin EMG during apnea/hypopnea onset. The OSA and CSA onset showed steeper drop in the chin EMG activity compared with the hypopnea onset (with p<0.001). Utilizing features from chin EMG and oxygen saturation, the classification of 30-sec sleeping epoch gave an accuracy of 85% with high specificity (accuracy of detecting the normal epochs) of 92%. Although the sensitivity (the accuracy of detecting apnea/hypopnea epochs) was 67%, the sensitivity of the OSA-only epochs was relatively high (85%). Further studies are needed to explore the reason for chin EMG drop during apnea onset.

Keywords: Cognitive Informatics, Biomarker Discovery, Machine Learning
Abstract: Recent developments in cognitive neuroscience have emphasized the use of naturalistic experimental paradigms, especially for real-world tasks like driving. This research introduced a blink-locked EEG segmentation method and contrasted its efficacy with traditional EEG segmentation. For three difficulty levels of proactive and reactive driving, we show a significant improvement in classification accuracy using a multi-classifier SVM with the blink-locked method, indicating enhancements of 4.3% for proactive driving and 4.4% for reactive driving. These findings underscore the potential of leveraging physiological markers, such as eye blinks, to enhance EEG data segmentation and deepen our understanding of cognitive dynamics in real-life scenarios.

Keywords: Deep Learning, Medical Imaging, Artificial Intelligence
Abstract: EEG signals frequently contain artifacts, which hinders analysis and requires time-intensive artifact removal. In this study, seeking to improve automation, we introduce an LSTM Autoencoder designed for EEG epoch classification through the anomaly detection approach. Its performance is benchmarked against two state-of-the-art denoising Convolutional Autoencoders. The implemented network exhibits high classification performance of clean and noisy epochs and additionally facilitates signal quality interpretation.

Keywords: Health Monitoring, Health Data Science, Digital Health
Abstract: This paper presents a possible application of mCrutch, a mobile health system consisting of a pair of sensorized crutches and a mobile Android application, with the purpose of recording the applied force and crutch orientation during a 6-minute walking test with a single crutch. We developed a MATLAB-based threshold algorithm for the segmentation of stance crutch events in analogy with gait analysis. A morphological analysis was performed to quantitatively and visually depict various stance features. The threshold algorithm achieved a segmentation accuracy of 94%. The whole application helps in gaining a comprehensive quantification of stance-related features across various phases of the test. Clinical Relevance— This approach facilitates the design and development of additional algorithms aimed at gathering clinically relevant parameters pertaining to crutch movement within the context of the human-crutch system.

Keywords: Medical Imaging, Neural Networks, Deep Learning
Abstract: Analytical solutions can be used to generate data to train Deep Learning neural networks to estimate electromagnetic fields. In this work, we compare the efficacy of two neural networks when two 2D analytical solutions are used to generate the training data: a geometry with 2 concentric infinitely long cylinders, and a geometry with up to 8 concentric infinitely long cylinders. The neural networks can estimate the electric fields and are trained with B1+ and SNR maps. The validation process was performed with results obtained with 3D numerical simulations. Even if more layers should provide higher heterogeneity in the training process, no significant improvement has been achieved with training with more layers, suggesting that it might be necessary to generate more data for better training with more heterogeneous geometries.

Keywords: Human–Computer Interaction, Machine Learning, Computational Biology
Abstract: Brain-computer interface (BCI) technology demonstrated immense potential across diverse fields. However, current research on electroencephalogram (EEG) commonly assumed single model can be universally applied across all gender identities. While a few studies have identified gender differences through supervised classification of resting-state EEG, the majority have largely overlooked the potential impact of gender-specific differences in BCI applications. This study explored the gender-specific differences in motor imagery (MI) within BCIs and the feasibility of gender recognition in unsupervised settings. We utilized public datasets of male and female EEG signals, applied widely used machine learning algorithms for task and gender identification. The results showed that the average MI classification accuracy for female was 0.57% higher than male, despite the dataset containing more male subjects. In addition, gender recognition accuracy from EEG MI data exceeded 97%. These findings have highlighted the importance of considering gender-specific differences in BCI research and application. The results of this study could inform the development of more personalized effective BCIs in healthcare and other fields, ultimately leading to improved outcomes and experiences for users of all genders.

Keywords: Health Data, Health Data Science, Healthcare Analytics
Abstract: Minimal-invasive surgery provides patients with benefits such as fewer incisions, faster healing times, reduced postoperative pain and bleeding, minimized scarring, and shorter hospital stays. To enhance its effectiveness ultrasound imaging or magnetic resonance-based guidance systems have been developed. However, research in robotics, sensors, and medical imaging is quickly evolving often presenting new solutions to the still persisting artifact and real-time feedback issues. To test those navigation technologies specialized phantoms are needed. Currently, existing phantoms are offering small tissue-like materials variability and are made of materials not suitable for extensive testing of minimally invasive surgery technologies. In this paper, we present a preliminary study about dedicated phantom design and manufacturing process, and data collection with special attention to vibroacoustic signals for future analysis.

Keywords: Health Monitoring, Machine Learning, Wearable Devices
Abstract: Acute psychological stress (APS) is a complex multifactorial event caused by drivers such as anxiety, mental and competition stress. It can occur concurrently with physical activity (PA), making its detection and classification challenging. This study investigates the detection and classification of APS (alone or concurrent with PA) by using physiological signals collected using Empatica E4 wristband. Multi-task Extreme Gradient Boosting (XGBoost) achieved F1 scores of 99.89% and 98.31% for the classification of various PA (treadmill run, stationary bike) and APS (competitive mental, anxiety stress, non-stress) and sedentary state. Shapley additive explanations (SHAP) is used to interpret the global importance of the physiological signals, determining the order of importance physiological signals for APS detection and classification. The results indicate (in decreasing order of importance): galvanic skin response (GSR), heart rate (HR), skin temperature (ST), accelerometer (ACC) X-axis, ACC Y-axis, ACC Z-axis. The increase in GSR and HR are positively correlated with the occurrence of APS as indicated by high positive SHAP values.

Keywords: Machine Learning, Forecasting, Prediction Models
Abstract: Traditional treatments do not work on 33% of epileptic patients.Warning devices employing seizure prediction or forecasting algorithms could bring patients a newfound quality of life. These algorithms would attempt to detect the preictal period, a transitional moment between regular brain activity and the seizure, and warn the user. Several past methodologies have been developed, triggering an alarm when detecting the preictal period, but few have been clinically applicable. Recent studies have suggested a paradigm change to seizure forecasting that takes a probabilistic approach instead of the crisp one of seizure prediction. The alarm is substituted by a constant risk assessment analysis. To the best of our knowledge, no direct comparison between prediction and forecasting using the same database has been made. This paper explores methodologies capable of seizure forecasting and compares them with seizure prediction ones. Using data from the EPILEPSIAE database, we developed several patient-specific prediction and forecasting algorithms with different classifiers (a Logistic Regression, a 15 Support Vector Machines ensemble, and a 15 Shallow Neural Networks ensemble). Results show an increase of the sensitivity in forecasting relative to prediction of up to 146% and in the number of patients displaying an improvement over chance of up to 200%. These results suggest that seizure forecasting may be more suitable for seizure warning devices than seizure prediction.

Keywords: Cognitive Computing, Machine Learning, Health Data Science
Abstract: We propose a completely data driven two pass electrode selection technique based on functional data analysis for EEG classification. In the first pass, electrodes are selected based on their activity and in the second pass, selected electrodes are retained based on their discriminant activity. Electrodes are ranked and based on their connectivity, informative clusters are extracted. We test our method on a face familiarity task and demonstrate the efficacy of our channel selection method over all channels and commonly used channels in face familiarity task.

Keywords: Computational Biology, Artificial Intelligence, Deep Learning
Abstract: With the soaring interest in understanding the dynamics of human body skeletons for applications such as action recognition and video understanding, the significance of precise 3D key-point detection has become increasingly promi- nent. Despite the advancements, existing approaches struggle to address the issues of occlusions and limited annotated data. This paper proposes a novel framework integrating a multi- level attention mechanism and weakly supervised 3D key-point generation to tackle these prevalent issues, enhancing both the accuracy and efficiency of human pose estimation.

Keywords: Natural Language Processing, Machine Learning, Informatics
Abstract: Social Networking Services (SNS), like Twitter (now ‘X’), hold promise for Blood Donation (BD)service enhancement. However, SNS data often contains noise, limiting its BD utility. We propose a solution employing Bidirectional Encoder Representation from Transformers (BERT) models combined with manual labeling. We collected Japanese BD-related tweets, creating the BD Tweet-User Classification (BDT-UC) dataset. BDT-UC includes Donor, Non-Donor, Undetermined, and optional Potential/Deferred labels, with English translations. We used BDT-UC to train and validate Japanese BERT models through two methods: CAT1-Models (fine-tuning) and CAT2-Models (customization with task-specific layers), considering data imbalance and cleaning for some tests. CAT1-Models served as a baseline, showing decent performance (up to 78% accuracy) despite data imbalance, especially in identifying Undetermined data (up to 85% Precision and 83.5% F1-Score). CAT2-Models showed similar consistent results regarding noise, even with 5 data categories. Our study offers a promising methodology for BD-related Twitter data classification and noise reduction. While focusing on Japanese tweets, we aim to make our Python program and dataset publicly available for broader language and region applications. In conclusion, the proposed tools shows promise to facilitate BD-related Twitter data analysis, enabling efficient categorization without information loss.

Keywords: Health Analytics, Clinical Data Science, Wearable Devices
Abstract: In this work, we describe the use of time series features extracted from foot temperature data obtained from a wearable in-shoe system to discriminate between feet from individuals affected by diabetes, and feet from healthy individuals. We identify a set of features that are statistically significant in discriminating between the two classes and that can thus can serve as input to machine learning classifiers.

Keywords: Artificial Intelligence, Cognitive Computing, Human–Computer Interaction
Abstract: Decoding visual representations from human brain activity has emerged as a thriving research domain, particularly in the context of brain-computer interfaces. This study introduces a novel approach using a convolutional neural network (CNN) to classify images from the ImageNet dataset, leveraging electroencephalography (EEG) recordings. We collected EEG data from 6 subjects, each viewing 50 images across 40 distinct semantic categories. These EEG signals were transformed into spectrograms, serving as the input for training our CNN. A unique aspect of our model is the incorporation of knowledge distillation from a pre-trained image classification teacher network. This approach enabled our model to achieve a top-5 accuracy of more than 80%, notably surpassing a plain CNN baseline. Furthermore, we integrated an image reconstruction pipeline founded on pre-trained latent diffusion models. This innovative concatenation not only decodes images from brain activity but also provides a plausible reconstruction, facilitating rapid and subject-specific feedback experiments. Our work thus represents a significant advancement in the field, bridging the gap between neural signals and visual perception.

Keywords: Big Data, Deep Learning, Healthcare
Abstract: Within the scope of prostate cancer diagnostic imaging, distinguishing lesions is challenging due to their subtle appearance and the prostate gland's complexity. This study employed the nnU-Net, a state-of-the-art medical image segmentation model, on a dataset of 301 patients from 2 openly available datasets to identify whether lesion sizes are affecting the model's performance. By applying t-SNE dimensionality reduction algorithm among dice score and respective lesion sizes, we found that nnU-Net behaves differently for lesions smaller than 9mm compared to lesions larger than 15mm. These insights can inform specialized training approaches for future deep learning models in prostate lesion detection.

Keywords: Artificial Intelligence, Data Mining, Data Analytics
Abstract: Nowadays, the existence of data silos obscures any advancements in healthcare. On the other hand, conventional centralized data analysis faces numerous obstacles, including legal issues, reduced levels of trustworthiness and lack of data interoperability. Federated artificial intelligence (AI) is an emerging strategy which enables collaborative model training across multiple centers without the need to centralize sensitive patient data. However, the biases which are introduced during the training process combined with the need to install equipment on premises remain underexplored. In this work, we propose a trustworthy, cloud-based AI framework, where federated implementations of high-performance boosting classifiers with hybrid loss functions were developed to solve supervised learning tasks and to provide interpretable risk factors. A case study was conducted to solve an unmet mucosa-associated lymphoid tissue (MALT) lymphoma classification problem by utilizing a PanEuropean data hub with 4805 patients with primary Sjogren’s Syndrome (pSS) (21 European cohorts). Our results highlight the performance of the federated boosting classifiers (0.9 AUC) along with explainable risk factors.

Keywords: Machine Learning, Health Data, Medical Imaging
Abstract: Coronary Artery Disease (CAD) represents a life-threatening condition resulting from the constriction or obstruction of coronary arteries. Timely detection plays a critical role in effective treatment. This paper introduces an innovative machine-learning approach that utilizes the ResNet architecture to automate stenosis identification in coronarography. The proposed method aims to enhance the efficiency and reliability of CAD diagnosis, providing valuable support to medical practitioners. Results emphasize the significance of high-quality training datasets in achieving precise stenosis detection. Discussion is made regarding study limitations, including dataset artifacts, and avenues for future research are proposed. This approach establishes a foundation for advancements in coronary vessel stenosis detection, with potential for classification and additional feature extraction.

The proposed method aims at the automation of CAD - a world-leading cause of death - diagnosis and personalized treatment suggestions. AI-supported detection and characterization facilitates cardiologist work required for manual data analysis.

Keywords: Deep Learning, Medical Imaging, Decision Support Systems
Abstract: Monitoring the changes in bone marrow during therapy for multiple myeloma patients is a crucial task. Osteolytic lesions can cause deformation of the bones, affecting the robustness of traditional segmentation tools. A two-model deep learning analysis is explored in this study. A detection model reduces pixel imbalances between the background and the bone marrow pixels, achieving a mAP of 0.878±0.005. A residual U-Net segments the bone marrow, yielding a DSC of 0.856±0.003. The proposed deep learning-based segmentation pipeline allows accurate and fast annotation of the bone marrow in multiple myeloma patients.

Keywords: Neural Networks, Computational Biology, Evidence-Based Healthcare
Abstract: Deep neural networks (DNNs) demonstrated excellent performance in fully-automatic interpretation of biomedical time series such as electrocardiography (ECG). Next to conventional clinical use-cases such as diagnosis of cardiac diseases, completely novel risk assessment methods emerged. For example, predicting age from raw ECG established as a novel measure of cardiovascular health with a greater ECG age than the chronological age being linked to higher mortality. However, due to the black box nature of DNNs, there is a lack in explainability, impeding adoption to clinical practice. In this work we aim to explain an open source DNN for age prediction which was trained on more than 1.5 million 12-lead ECG recordings acquired in Brazil. First, we perform a validation study of the model by predicting a German dataset (PTB-XL) and observe similar results indicating generalization of the model over different patient populations and health care systems. Second, we conduct an ablation study by randomly masking single ECG leads for every patient to analyze the model's performance w.r.t lead importance using Pearson correlation coefficient. Our analysis indicates that leads I and V1 have the highest influence on age prediction. With a drop of model performance to 0.93 and 0.90, respectively, this might indicate the importance of the atrial depolarization.

Our results might strengthen the trust in DNN-based methods and ECG-based age prediction.

Keywords: Drug Discovery, Neural Networks, Machine Learning
Abstract: Recent advances in deep learning have enabled the screening of drugs from large datasets, the study of new compounds, and the interpretation of the significance of substructures with target characteristics. In this work, we explored in silico analysis by utilizing a convolutional neural network and a transformer in modeling the efficacy of molecules ranging from weak to strong in their binding with sodium ions as an example. We adopted the most representative models, including the VGG19 model and the vision transformer (ViT), trained them on 2D depictions of molecular structures to predict their equilibrium constants, and compared the predictions to the experimental results. VGG19 and ViT both achieved a normalized RMSE of 9% and 8%, respectively, in predicting the equilibrium constant, while also achieving Pearson Correlation Coefficients of 0.80 and 0.83, respectively, between the predicted and experimental results data. This demonstrates the great predictive capacity of well-established computer vision models for this task. Clinical Relevance — This work predicts the binding constants between a large number (>1000) of molecules and sodium ions. Identification of key structural properties of the molecules can increase the hit rate in primary drug screening assays.

Keywords: Wearable Devices, Digital Health, Machine Learning
Abstract: Reliable and thorough pain assessment is essential for effective pain treatment, starting with the correct identification of the type of pain, the two most common being nociceptive and neuropathic pain. We aim to train machine learning algorithms, fed with data from wearable sensors, to be used as a diagnostic differentiation tool to distinguish between nociceptive and neuropathic pain experienced by patients undergoing neurorehabilitation.

Keywords: Digital Health, Decision Support Systems, Artificial Intelligence
Abstract: This study aimed to propose an approach that can be used to predict the risk of sarcopenia in non-clinical settings. A total of 90 participants were divided into three study groups: 30 participants with low-sarcopenic risk, 30 with mid-sarcopenic risk, and 30 with high-sarcopenic risk. Each participant was instructed to sit on a chair, and a device equipped with an electrical stimulator, a surface electromyogram acquisition module, and an electrode module was attached to the skin surface on the rectus femoris and biceps femoris muscles of the dominant leg. While delivering multi-frequency electrical stimulation to the muscles, the device measured muscle response signals. Ten parameters to quantify nonlinearity in time-series data were time-sequentially extracted from the measured signals. The sequences of the parameters were fed into a bidirectional long short-term memory layer, and then the output was integrated with demographic characteristics in a feature fusion layer followed by a fully-connected layer. The 80% of participants in each study group were used for classifier training and validation, and the remaining 20% of participants in each study group were used to test the classifier. The classifier achieved a test accuracy of 0.944 in classifying the low-, mid-, and high-sarcopenic risk groups. This study would pave the way for in-home self-monitoring of sarcopenic risk, which can contribute to early and effective prevention of sarcopenia.

Keywords: Cognitive Computing, Cognitive Informatics, Wearable Devices
Abstract: This study examined the mechanisms that enhance the overall accuracy of mental workload assessment by synergizing entropy within and between EEG channels. First, we filtered the EEG signals recorded from prefrontal channels and divided them into sub-bands. Next, we derived cross-approximate entropy (XApEn) and conventional approximate entropy (ApEn) metrics from each sub-band. Finally, these derived features were input into an AdaBoost classifier for mental workload assessment. Comparing the classification results reveals that the combination of XApEn and ApEn outperformed the separate employment of XApEn or ApEn, achieving a higher mean accuracy of 82% vs. 72% and 70%, respectively. The results show that interaction between the prefrontal EEG channels is important when assessing mental workload.

Keywords: Personalized Healthcare, Artificial Intelligence, Bioinformatics
Abstract: Facial expressions, speech, and physiological signals are primarily used for human emotion recognition. Physiological signals are controlled by the autonomic nervous system. These physiological signals can vary individually in response to the same physical stimuli which results to lower accuracy in recognizing emotions when compared with other modalities. In this paper, emotion recognition accuracy is enhanced by considering personality traits. To achieve this, a new dataset is collected which includes physiological signals mapped with different personality types used to train an emotion recognition system. The accuracy of emotion recognition was compared based on personality types indicating judgment functions.

Keywords: Health Data, Machine Learning, Healthcare
Abstract: Abstract—Major challenge in building models for stroke rehabilitation stems from the non stationarity of the EMG signals. In this work we present a methodology for improved classification of hand gestures using Empirical Fourier Decomposition (EFD). First we apply the EFD technique on a set of publicly available dataset and later we reduce the dimensionality to collect most significant components. Finally we extract features from these components and perform hand gesture classifications using different machine learning (ML) models.

Clinical Relevance—Compared to the state-of-art Empirical Wavelet Transform (EWT), the EFD technique reduced the total significant components considerably. To capture 90% of information from original data, the EFD approach needed 5.96% and 23.21% less number of components compared to EWT approach for original and dimensionally reduced data sets respectively. The classification models using EFD components gave an average 3.4% accuracy improvement compared to that of EWT components.

Keywords: Biomarker Discovery, Clustering , Precision Medicine
Abstract: Chronic pain is a widely prevalent and difficult to treat disease due to the heterogeneity of the patient experience and lack of objective biomarkers to characterize its physiology. There has been recent interest in a specific brain oscillation (alpha) measurable with electroencephalography (EEG) as potentially linked to the specific phenotype of chronic pain. In this study, we investigate a fully data-driven approach to identifying the exact frequency of this alpha oscillation and characterizing its spatial variation across the scalp in a patient-specific manner. We also cluster patients based on this spatial signature using unsupervised methods and show that this clustering has potential for clinical relevance based on self-reported pain scores.

Clinical Relevance — Objective, data-driven stratification of chronic pain patients using EEG enables personalized disease management and physiologic insight into varied disease etiology.

Keywords: Neural Networks, Data Science, Data Analytics
Abstract: Abstract— High quality feature sets are vital to the development of neuroprosthetic and bioelectronic therapies. For neural signals, the primary features are “spikes” extracted from neuronal signals. Current spike detection is a subjective process with no certain ground truth. Leveraging synthetic neuronal data, we can train a deep learning model to provide continuous pointwise spike probability and leverage this to improve feature detection quality. Our results indicate that high-fidelity, variable-length spikes can be extracted using continuous probability

Clinical Relevance— This work improves feature quality and diversity over current static thresholding techniques. Furthermore, focusing on small-amplitude neuronal signals gives vastly superior insight into autonomic regulatory encoding. This is particularly important for the development of bioelectronics pain treatment and synthetic organ development, as the smaller-amplitude longer-time features they need are frequently missed by current approaches.

Keywords: Medical Imaging, Medical Diagnostics, Artificial Intelligence
Abstract: Crohn's disease (CD), a chronic inflammatory bowel disorder, often affects the terminal ileum (TI) and leads to digestive tract inflammation and complications like bowel obstruction. Accurately determining the 3D extent of CD from 2D Magnetic Resonance Enterography (MRE) images requires approximations, as no automated 3D measurement system exists. We developed an intelligent MRE reading application for virtual unfolding and 3D visualization of MRE data. We introduce a semi-automatic algorithm to predict the TI centerline from MRE data to reduce radiologist interaction time. The algorithm involves constructing an orientation classifier and implementing a shortest path algorithm to determine the TI centerline. We evaluated the algorithm's effectiveness on a database of 123 MRE scans using a k-fold cross-validation experimental setup, comparing the predicted centerline with a radiologist-annotated centerline considered as ground-truth. The results showed good alignment with the ground-truth centerline with minimal interaction time.

Keywords: Deep Learning, Medical Imaging, Radiomics
Abstract: Predicting pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) in breast cancer is crucial for effective surgical planning and treatment optimization. While radiomics-based methods have been explored for pCR prediction using diffusion-weighted MRI (DWI), they rely on manual tumor segmentation—a laborious and error-prone task. Our study introduces a deep learning model that automates tumor segmentation from DWI, enhancing the accuracy and efficiency of radiomics-based pCR predictions and eliminating the need for manual intervention. We evaluated our approach on the publicly BMMR2 challenge data using a k-fold cross-validation experimental setup, comparing the radiomics-based pCR predictions from manual and automatic segmentations. Our approach demonstrated a human-level performance for pre-treatment radiomics-based pCR prediction from the DWI data.

Keywords: Deep Learning, Medical Imaging, Big Data Analytics
Abstract: Federated Learning (FL) enables collaborative model training among medical centers without sharing private data. However, traditional FL risks on server failures and suboptimal performance on local data due to the nature of centralized model aggregation. To address these issues, we present Gossip Mutual Learning (GML), a decentralized framework that uses Gossip Protocol for direct peer-to-peer communication. In addition, GML encourages each site to optimize its local model through mutual learning to account for data variations among different sites. For the task of tumor segmentation using 146 cases from four clinical sites in BraTS 2021 dataset, we demonstrated GML outperformed local models and achieved similar performance as FedAvg with only 25% communication overhead

Keywords: Medical Imaging, Deep Learning, Machine Learning
Abstract: Scattered light is unavoidable during biomedical imaging, leading to downgraded images. Removing such scattered light, however, is challenging, because the imaging objects are complex and the light path is complicated. In this paper, we propose to use deep learning methods to computationally remove such scattered light. Inspired by the dehazing methods for natural images, we first developed a novel mathematical model for biomedical images. Then, we developed a deep neural network method to solve this model. Our network simultaneously estimates the clear image and the scattered light. Several experiments are conducted to confirm the network's effectiveness and efficiency.

Keywords: Artificial Intelligence, Decision Support, Deep Learning
Abstract: In large hospitals, multiple radiologists share the imaging workload. In collaborative work, each radiologist's work goal is often set as the number of cases, but even with the same number of cases, there can be an uneven workload. If radiologists select images that seem less workload based on the order content and leave images that seem more workload, images that should have been diagnosed earlier are delayed. To prevent this, some readers give priority to tests that seem emergent or have a long time passed from the time of imaging, creating a sense of unfairness among radiologists. Thus, it may be possible to reduce unfairness by labeling each image with an indicator of workload and distributing or revealing the work in advance. In this study, we predict workload indices for each image based on order content and patient information, without using images, for CT images used in collaborative diagnostic work. The dataset consisted of CT image order content and patient information at Kyoto University Hospital. Next, a radiologist referred to order content, and 1097 samples were labeled in three levels according to their diagnostic workload. We considered a total of four pre-trained models and experimented with multimodal datasets. Training used the remaining data for five-fold cross-validation to train the proposed model. The proposed model has a macro-F1 0.682, and the confusion matrix shows a distribution in prediction.

Keywords: Deep Learning, Medical Imaging, Medical Diagnostics
Abstract: In this research we aim to develop a Computer-aided diagnostic system(CAD) for Breast Ultrasound Imaging to enhance the early detection of breast cancer. CAD tools aid physicians in the diagnostic process leading to early detection and treatment of breast cancer. The proposed CAD tool comprises three stages, the first is Enhancement of breast ultrasound imaging by applying custom segmentation deep learning architectures[U-Net, SK-U-Net, RDA-U-Net] to enhance the isolation of masses in Ultrasound images, then Generation of new ultrasound images by employing Generative Adversarial Networks[DAGAN, DCGAN] to overcome data limitation in available data-sets, and the last stage is the Detection and Classification of breast cancer masses by a CNN model. The Enhancement stage shows promising results with accuracy of 98 % overall accuracy, 92 % overall precision, and 85 % overall sensitivity.