Over the past five years, artificial intelligence (AI) has introduced new models and methods for addressing the challenges associated with the broader adoption of AI models and systems in medicine. This paper reviews recent advances in AI for medical image and video analysis, outlines emerging paradigms, highlights pathways for successful clinical translation, and provides recommendations for future work. Hybrid Convolutional Neural Network (CNN) Transformer architectures now deliver state-of-the-art results in segmentation, classification, reconstruction, synthesis, and registration. Foundation and generative AI models enable the use of transfer learning to smaller datasets with limited ground truth. Federated learning supports privacy-preserving collaboration across institutions. Explainable and trustworthy AI approaches have become essential to foster clinician trust, ensure regulatory compliance, and facilitate ethical deployment. Together, these developments pave the way for integrating AI into radiology, pathology, and wider healthcare workflows.

This study evaluates the effectiveness of various non-echocardiographic medical biomarkers, integrated with machine learning (ML) models, for classification of different types of Heart Failure (HF), specifically heart failure with preserved ejection fraction (HFpEF) and heart failure with reduced ejection fraction (HFrEF). This differentiation is critical due to the distinct pathophysiology and management strategies required for these heart failure subtypes. A retrospective clinical dataset was collected from three medical centers in Serbia, comprising anonymized records of 481 heart failure patients ($221 \text{HFrEF}, 260 \text{HFpEF}$). The dataset included three primary types of variables: health records, laboratory test results, and electrocardiogram (ECG) data. Data preprocessing involved standardization of laboratory values and Multivariate Imputation by Chained Equations (MICE) to address missing parameters. Five different machine learning algorithms were employed, with an$80 / 20$train-test split and exhaustive Grid Search for hyperparameter optimization: Decision Tree, Random Forest, eXtreme Gradient Boosting Tree (XGBoost), Multilayer Perceptron (MLP), and Gradient Boosting Tree. Model performance was assessed using accuracy, precision, recall, and$\mathbf{F 1}$-score. The results indicate that ensemble methods, specifically Gradient Boosting Tree and Random Forest, consistently achieved superior performance across various data subsets. Notably, the combination of Health Records and ECG data yielded the highest predictive performance, with the Gradient Boosting Tree model achieving an F1-score of 0.7736 and a recall of 0.8039. Conversely, models trained predominantly on, or solely with, laboratory testing features consistently exhibited lower performance, suggesting that a broad inclusion of these biomarkers may introduce noise or redundancy for this specific classification task. These findings demonstrate that non-echocardiographic biomarkers can be effectively leveraged by machine learning models to differentiate between HFrEF and HFpEF, offering a promising diagnostic tool.

Heart failure (HF) is one of the most common medical conditions around the world in the modern age. Patients with HF struggle with reduced quality of life and have a greater risk of death. For these reasons it is important for HF to be diagnosed early before it starts to majorly impact patients and cause major concern for death. Left ventricular ejection fraction (LVEF) is one of the most important features which can be used to further classify HF patients into patients with reduced ejection fraction (HFrEF), mid-range ejection fraction (HFmrEF) and preserved ejection fraction (HFpEF). In this paper we describe a machine learning based pipeline for multistage classification of patients according to presence of HF and degree of reduced ejection fraction. Two separate pipelines are created in the scope of our research. Both pipelines classify patients based on whether HF is present or not. The first pipeline classifies HF patients into 3 classes based on their ejection fraction into reduced, mid-range and preserved classes. The second pipeline conducts classification of patients into LVEF$<50 {\%}$and LVEF$>50 {\%}$based on available features, followed by a second classification of HFrEF and HFpEF classes. HF classification model achieved 97% accuracy and$\mathbf{9 8 \%} \mathbf{F 1}$score for the confirmed HF class. The 3-class classification model achieved an overall prediction accuracy of 82% with F1 scores of 90%, 51% and 86% for HFpEF, HFmrEF and HFrEF respectively. The final two stage classification model achieved an overall accuracy of 96% and 87% for LVEF$>50$% versus LVEF$<50$% and mid-range versus reduced ejection fraction respectively. In the case of the twostage classification model F1 scores of$\mathbf{9 2 \%, 9 5 \%}$and 97% were achieved for HFpEF, HFmrEF and HFrEF respectively.

Image segmentation is the process of assigning a label to every pixel in an image such that pixels with the same label share certain characteristics. Traditionally image segmentation was used on 2D imaging data, but due to the increased usage of 3D volumetric data there is a need for 3D segmentation techniques that could utilize spatial information contained in these volumes. One of the fields where there is a great amount of 3D data is microscopy. This paper introduces convolutional neural network based on 3D U-net architecture for segmentation of confocal microscopy images of cells in an in vitro sprouting angiogenesis model. Developed model contains 4 layers where each encoder block contains two 3D convolutional layers, Batch Normalization, ReLU activation function and 3D max pooling layer, while each decoder block contains upconvolution, skip connections and two 3D convolutional layers. Preprocessing of this data resulted in the volumes of shape 256 × 256 × 256 voxels which were used for training of the developed model. The model achieves great segmentation results as showed by Jaccard index value of 94.52% and Dice coefficient value of 99.31% compared to the preprocessed dataset. Even when segmentation results are compared to the original dataset, model still achieves respectable results of 84.22% Jaccard index and 88.18% Dice coefficient. This introduction of automatic 3D image segmentation could greatly reduce the time required for data preparation, while achieving high degree of segmentation accuracy.

The use of image segmentation is widespread in many different fields of research including medicine, biology, computer vision and others. In most cases, image segmentation is done through the use of supervised learning approaches, which utilize ground truth segmentation masks. However, in biomedical research, segmentation masks are often not available, which leads to the use of unsupervised segmentation approaches. In these situations, filtering paired with region expansion and reduction and edge detection techniques are used, which in some cases require time-consuming manual parameter tuning, in order to achieve satisfactory results. This paper implements fully unsupervised image segmentation based on a deep learning convolutional neural network to segment cell geometries from confocal microscopy images of an in vitro model of angiogenesis. The proposed network architecture is a W-Net which contains a single U-Encoder and U-Decoder, both of which contain 3 convolution blocks and 3 deconvolution blocks, with 2 convolution layers and one max-pooling or up-convolution layer respectively. The results were evaluated using available ground truth images and achieved a pixel wise classification accuracy score of 98.73% and 73.71% intersection over union. The achieved high degree of accuracy shows great promise in cell segmentation without a need for ground truth masks over time and increases the accuracy of calculations of exerted forces.

Heart failure is recognized as a modern epidemic and despite advances in therapy and research, heart failure still carries an ominous prognosis and a significant socioeconomic burden. The main aim of this paper is to demonstrate how novel Decision Support System (DSS) and computational platform like INTELHEART can transform the future of healthcare and early diagnosis of heart failure. The main idea is integration of patient-specific data (i.e. demographic and physical characteristics, medical history, symptoms and signs) and results obtained using existing and novel diagnostic technologies into the cloud environment. Data will be used by different tools for machine learning and computational modelling, developing virtual patient population. Moreover, voice as a biomarker will be collected among participating patients, in order to create a VoiceHeart mobile app. INTELHEART represents a transformative advancement in heart failure care, aiming to make treatment more personalized, and proactive. This initiative centers on precision medicine, using AI-driven analysis and a powerful DSS alongside the cloud-based platform and VoiceHeart mobile app to assist both clinicians and patients. Additionally, it incorporates assessments of psychological resilience and emotional well-being, addressing the oftenoverlooked mental health factors essential to comprehensive heart failure management.

Medical imaging plays an important role in medicine today, assisting in illness diagnosis and therapy. For limited medical image datasets, training from scratch is not an option, hence transfer learning emerges as a solution, with ImageNet weights being utilized as initial weights, followed by fine-tuning. This paper takes a different approach by introducing transfer learning approach with pretrained architecture DenseNet121 with CheXNeXt weights. Collected dataset consisted of 227269 X-ray images from public databases and 684 chest X-ray images from a retrospective study conducted in the University Clinical Center of Kragujevac and includes information on atelectasis, cardiomegaly, parenchymal consolidation, edema, effusion, emphysema, fibrosis, hiatus hernia, infiltration, pleural thickening, non-viral pneumonia, pneumothorax, viral pneumonia in the form of Covid-19, tuberculosis as well as tumors in the form of mass and nodules. The results show that the model is able to distinguish between the healthy and diseased lungs with average AUC of 0.91 (the lowest AUC of 0.8 for emphysema and the highest AUC for of 0.99 for pneumonia and 0.98 for COVID-19). Although the results seem promising, additional fine tuning may be necessary to improve other metrics. Future research will focus on this aspect, as well as on creating a glass box system for classification.

Localization of lumbar discs in magnetic resonance imaging (MRI) is a challenging task, due to a vast range of shape, size, number, and appearance of discs and vertebrae. Based on a review of the cutting-edge methods, the majority of applied techniques are either semi-automatic, extremely sensitive to change in parameters, or involve further modification of the results. All of the above represents a motivation for implementing deep learning-based approaches for automatic segmentation and classification of disc herniation in MR images. This paper proposes a complete automated process based on deep learning to diagnose disc herniation. The methodology includes several steps starting from segmentation of region of interest (ROI), in this case disc area, bounding box cropping and enhancement of ROI, after which the image is classified based on convolutional neural network (CNN) into adequate classes (healthy, bulge, central, right or left herniation for axial view and healthy, L4/L5, L5/S1 level of herniation in sagittal view). The results show high accuracy of segmentation for both axial view (dice = 0.961, IOU = 0.925) and sagittal view (dice = 0.897, IOU = 0.813) images. After cropping and enhancing the region of interest, accuracy of classification was 0.87 for axial view images and 0.91 for sagittal view images. Comparison with the literature shows that proposed methodology outperforms state-of-the-art results when it comes to multiclassification problems. A fully automated decision support system for disc hernia diagnosis can assist in generating diagnostic findings in a timely manner, while human mistakes caused by cognitive overload and procedure-related errors can be reduced.

Since the novel SARS-CoV-2 virus appeared, interest in developing epidemiological mechanisms that would help in prevention of its spread has increased. Epidemiological models are the most important mechanisms for examining the spread of the virus. For that purpose, we propose deep learning approach, LSTM neural network model. LSTM is a special kind of neural network structure capable of learning long-term dependencies in sequence prediction problems. The model was fed with official statistical data available online for Belgium in the period of March 15th, 2020 to March 15th, 2021. Results show that LSTM is capable of predicting in long-term manner with the low values of RMSE and MAE. Higher values of RMSE and MAE are observed in the infected cases (RMSE was 397.23 and MAE was 315.35) which is expected due to thousands of infected people per day in Belgium. In future studies, we will include more phenomena, especially medical intervention and asymptomatic infection, in order to better describe the COVID-19 spread and development.

Emitted fraction (EF) is one of the critical quality attributes of dry powders for inhalation (DPIs). Traditionally, different in vitro methods have been used for the assessment of DPIs EF. However, the evolution in computer-based (in silico) methods led to the development of special fields, such as Computational Fluid Dynamics (CFD) coupled with fluid-particle dynamics models e.g., Discrete Phase Model (DPM) as a useful alternative for the assessment of DPIs aerodynamic performance. The aim of this study was to design a CFD-DPM model for the prediction of model DPIs EF, and assess the prediction power of this method by comparing the in silico prediction results with in vitro determined EF values, obtained by three different methods. The EFs of the solid lipid microparticles (SLM) DPIs were determined in vitro by Twin stage impinger, Next generation impactor and Fast Screening Impactor. CFD-DPM model was successfully designed, and then the simulation results indicated the percentage of particles that remained in the inhaler. Based on those data, DPI EFs were calculated to be in the range of 83 % and 92%, indicating that CFD-DPM simulations were able to catch the differences between five SLM DPI formulations. In addition, CFD-DPM predicted the regional particle deposition in the inhaler, which cannot be precisely determined based on in vitro experiments. CFD-DPM predicted EF values were generally comparable to the EF values obtained by three in vitro methods, although some differences were observed between in vitro and in silico values. Therefore, it can be concluded that although additional improvements of CFD-DPM designed model are still necessary in order to be able to precisely describe aerodynamic performance of SLM DPIs, CFD-DPM modeling can be considered as a very useful tool in DPIs development.

The paper describes mathematical model and numerical simulation of Mason-Weaver equation using finite difference method, FDM, for simulation of sedimentation process. Different FDM schemes have been developed and tested for several different initial conditions. Possible issues with numerical convergence and conservation of concentration are explained. Performed analysis can be important for any numerical simulation that captures sedimentation process. The results of this research can be further used in modelling epithelial cell behavior and lung-on-a-chip systems.

Diagnosing lumbar discus hernia is a challenging task, due to disc and vertebral variations in size, shape, quantity, and appearance. Medical history and physical examination, electrodiagnostic tests, and MRIs are all used by doctors to set a definitive diagnosis. A majority of the state-of-the-art methods are semi-automatic and require extra corrections to the solution or are extremely sensitive to changes in parameters. Based on literature review, there is a solid basis for implementation of machine learning-based methods for disc herniation detection in MRI images. An automated segmentation method of vertebrae and discs is proposed in this study as a first step towards a decision support system for discus hernia identification. Dataset consisted of 104 images in sagittal and 99 images in axial views. Optimized convolutional neural network U-net has demonstrated very high accuracy in segmentation. Additional result represents the calculated distance from the disc's center to the disc's edge points in axial images across 360°, which results in clearly different number of peaks for the healthy and diseased discs. Fully automated computer diagnostic system helps speed up the process of setting up adequate diagnosis and reducing human mistakes.

The aim of this research was to analyze objectively the process of disc herniation identification using Bayes Theorem. One of the symptoms of discus hernia is muscle weakness on the foot that is caused by displaced discs in the space of two vertebrae. This fact is used by experts in initial diagnosis of herniated discs and we used it to create non-invasive platform for the same purposes by measuring force values from four sensors placed on both feet (first, second, and fourth metatarsal head as well as the heel). Dataset consisted of several minute force recordings of 56 subjects with discus hernia and 15 healthy individuals during normal standing, standing on forefeet and heels. The subjects were diagnosed by a specialist with either L4/L5 or L5/S1 discus hernia. Collected recordings were processed in several steps including filtering, extraction of forefeet and heel recordings, classification of average values for forefeet, and heel sensors to the groups with or without foot muscle weakness. Application of Bayes Theorem on the attributes of interest showed average 78.3% accuracy with 62.6% sensitivity and 80.9% specificity, while application of naive Bayes Network showed average 83.1% accuracy with 57.6% sensitivity and 88.2% specificity. Very weak or no correlation was observed between gender and disc hernia diagnosis (or obesity type and disc hernia diagnosis). Obtained results show that this method can be used in initial screening of patients and be a supportive tool to doctors to send the same patients for further examination.

The main aim of this study is to develop an application which will be able to load DICOM images, segment a tumor on brain slices and create a three-dimensional model of the segmented tumor. Firstly, we present two algorithms - depth first search method (DFS) and active contours method, for brain tumor segmentation and compare their performance. Comparison of effectiveness of these two proposed algorithms included discussing parameters complexity of initial conditions that need to be set manually, accuracy of the segmented tumor surface area, and computational time. Described methodology was tested on computerized tomography (CT) medical images from 37 patients using two different algorithms. User-friendly application for tumor segmentation on two-dimensional images and three-dimensional visualization is developed in Java. Presented application and results can be used as assistance tools in cases of surgeries, but can also be helpful in brain tumor treatment.

Hip replacement surgery is one of the most common and most successfully performed surgeries. Hip is an important joint in the human body that provides us with ability to perform different daily activities (walking, running, etc.). In this paper we have analyzed biomechanics of femoral bone with cementless hip prosthesis. The goal was to analyze stress distribution of the implant and femur bone. For numerical calculations of the stress distribution we have used finite element analysis. Presented results include von Mises stress distribution and Maximum Principal Stress distribution.

In this paper we present a method for brain tumor segmentation and assess its performance discussing parameters - complexity of initial conditions that need to be set manually, tumor surface area recognition, and computational time. The methodology includes performing segmentation on computerized tomography (CT) medical images from 37 patients. Furthermore, one approach to user friendly two- and three-dimensional tumor visualization is proposed. The results obtained in this paper can be new paradigm in the assessment of optimal approach trajectory to brain tumor in surgical operation.