The Latest Developments and Challenges in Biomedical Engineering

This paper considers the problem of corneal endothelium image segmentation using a method that combines a CNN model with a watershed transform. Specifically, first CNN predicts cell bodies, edges, and centers. Next, cell centers are used as markers that guide the watershed transform performed concerning the cell edge probability maps inferred by the CNN to outline cell edges. Different variants of the method are considered. Specifically, a downscaled U-Net is compared with the Attention U-Net in the image-to-image and sliding window setup. Results show that using a marker-driven watershed transform to post-process cell edge probability maps allows for replacing the sliding window setup with an image-to-image setup, reducing prediction time while maintaining similar or better segmentation accuracy. Also, when used as a backbone, Attention U-Net outperforms classical U-Net in determining cell morphometric parameters with high accuracy.

Tissue pattern is an important factor in morphological evaluation of tissue samples. It can be decisive in disease discrimination or for establishing disease subtypes. Tissue architecture can be generally described by pathologist as classical, hypocellular or hypercellular. This article presents a study used to establish classification convolutional neural network model for tissue compactness assessment. The VGG16 network was trained to classify image patches to create reliable heatmaps. The application of image augmentation, class-specific sampling, and hyperparameter tuning was used to prevent overfitting and increase the accuracy of the model. Based on the current results it can be concluded that the differentiation between hypo- and hypercellular tissue (compactness) is possible with application of deep learning classification model VGG16. We hope to find correlation of features related to tissue compactness with subtypes of the analysed disease that would become diagnostic markers or prognostic factors.

Background segmentation is an important step in analysis of histopathological images. It allows one to remove irrelevant regions and focus on the tissue of interest. However, background segmentation is challenging due to the variability of stain colors and intensity levels across different images, modalities, and magnification levels. In this paper, we present a learning-based model for histopathology background segmentation based on convolutional neural networks. We compare two multiresolution approaches to deal with the variability of magnification in histopathology images: (i) model that uses upscaling of smaller patches of the image, and (ii) model simultaneously trained on multiple resolution levels. Our model is characterized by solid performance both in multiresolution and multistain dyes (H &E and IHC), achieving good performance on publicly available dataset. The quantitative scores are, in terms of the Dice score, close to 94.71. The qualitative analysis presents strong performance on previously unseen cases from different distributions and various dyes. We freely release the model, weights, and ground-truth annotations to promote the open science and reproducible research.

Lymphocytes, a type of leukocytes, play a vital role in the immune system. The precise quantification, spatial arrangement and phenotypic characterization of lymphocytes within haematological or histopathological images can serve as a diagnostic indicator of a particular lesion. Artificial neural networks, employed for the detection of lymphocytes, not only can provide support to the work of histopathologists but also enable better disease monitoring and faster analysis of the general immune system condition. In this study, the impact of visual quality on the performance of state-of-the-art algorithms for detecting lymphocytes in medical images was examined. Two datasets were used, and image modifications such as blur, sharpness, brightness, and contrast were applied to assess the performance of YOLOv5 and RetinaNet models. The study revealed that the visual quality of images exerts a substantial impact on the effectiveness of the deep learning methods in detecting lymphocytes accurately. These findings have significant implications for deep learning approaches used in digital pathology.

A lot of time and effort has been put into finding an automatic segmentation system which could detect the region and position of white matter hypersensitivities in MRI brain scans. At the cellular level, changes in the white matter of the brain can be understood as a loss of myelin around axons. These changes might be detected by MRI due to local changes in water content. By establishing a method which could detect such irregularities, it would be possible to help diagnose some serious autoimmune diseases. In this work, a wide variety of local thresholding algorithms that are applied after local normalization is evaluated in terms of how accurately they segment these hypersensitivities. With the use of ImageJ, the following algorithms were tested: Bernsen, Contrast, Mean, Median, MidGrey, Otsu, Sauvola and SDA. The most accurate segmentation results are presented for all algorithms. For most of the ALT algorithms, the use of local normalization increased the achieved Dice value by as much as 100%. Contrary to what might be expected, the best results were achieved for various values of the sigma1 and sigma2 parameters of local normalization, both for single images and for a group of algorithms for a given image, thus it is difficult to apply the presented solutions to other examples.

White matter hyperintensities are regions of hyperintensive signal in white brain matter that appear in T2 or FLAIR MRI imaging. They show demyelination, cerebral oedema, tumors, or angiogenesis. Their presence is associated with a number of neurological diseases such as dementia, cognitive impairment, depression, schizophrenia, or even cerebral small vessel disease and multiple sclerosis. WMH areas are thought to appear on MRI images even a few years before the clinical manifestation of certain neurological diseases. Automatic segmentation algorithms can help better understand white matter lesions as part of large-scale studies. It is believed that changes in lesions’ appearance can be monitored over time, thus their correlation with certain diseases can be better understood. This work describes the use of histogram features for the detection of white matter hyperintensities, with a focus on skewness and kurtosis. The first aim of this study is to find features that can be used to identify white matter hyperintensities; then, the authors propose a fully automatic detection algorithm based on a sliding window method to obtain local histograms, and a Support Vector Machine algorithm to perform binary classification. The authors explored the effects of different combinations of characteristics and the relationship between feature values and hyperintensity percentage in a disc-shaped window. The described study is preliminary and contains an initial examination of the effectiveness of using histogram features for the detection of white matter hyperintensities in MRI images.

In this paper, we compare two approaches to automatically classify MR images of kidney lesions. The first approach involves the extraction of texture features for manually delineated kidney regions of interest (ROI) and then the classification of feature vectors with a model trained in a supervised manner. This classic machine learning approach is then challenged by a convolutional neural network-based method, which performs image classification in the learned latent feature space. In both cases, We aim to verify the hypothesis that it is possible to differentiate the state of renal failure between three classes: control, active inflammation, and chronic malformations based on the information content of the T1-weighted non-contrast enhanced MRI. The experiments performed on a sample of 25 showed superior performance of the convolutional neural network, for which we obtained the accuracy score at the level of 94% against 80% for the texture-based classification.

The increasing awareness and popularization of sign language in many countries have contributed to the development of mobile educational applications. This also applies to Polish Sign Language, used by the Deaf community in Poland. However, among the available solutions, there is a need for one which would engage the user in independent signing, i.e., forming the necessary finger motor skills and dexterity. To address this challenge, the paper focuses on developing a system for recognizing Polish Sign Language symbols that could work inside a mobile application. The publicly available MediaPipe library is used to determine the hand’s characteristic points and their coordinates. The detected points are then normalized and stored. Three machine learning techniques are verified at the character recognition stage: support vector machine, k-nearest neighbors, and random forest. Each was taught on the coordinate values obtained from a set containing 29,945 images of hands, divided into 29 classes (considering selected letters, supporting characters, and numbers). The obtained accuracies of all the methods are equal to about $$98\%$$ 98 % .

Breast cancer is a major global health concern affecting millions of women each year. Computer-aided diagnosis (CAD) systems have the potential to contribute significantly to early detection and reducing the mortality rate of breast cancer. This paper proposes a new methodology for breast cancer detection utilising data analytics, artificial intelligence, and mammograms. The approach is a mixed methodology based on colour clustering and deep transfer learning techniques to extract features from mammogram images. The proposed method was validated using the mini-DDSM mammogram images dataset, and its effectiveness was evaluated using various metrics such as accuracy, specificity, precision, recall, and F1 score. The results showed that all networks had high detection accuracy, with GoogleNet achieving the highest (99.58%) and ShuffleNet the lowest (97.08%). The proposed method achieved 100% detection accuracy using ResNet18, VGG16, ShuffleNet, DarkNet, and NasnetLarge, while Inception-ResNet-v2 had a detection accuracy of 98.33% with LRC and 99.17% with SVM. The proposed method has demonstrated the potential to improve the performance of CAD systems.

Panoramic radiograph images are one of the basic examinations used during dental and orthodontic treatment. Their creation is based on ionizing radiation, to which the patient is therefore exposed. The adverse effects of this radiation on the patient’s health is a topic that has been increasingly raised in numerous publications and journals. Therefore, the motivation for this study was to develop a mechanism that reconstructs panoramic radiograph images from MRI data, which is a harmless method. To achieve this, an algorithm described in a previous article is used to create a three-dimensional bone and skin tissue model from MRI data derived from T1-weighted and T2-weighted sequences. Subsequently, the craniofacial part of this model is projected onto a plane to form a panoramic radiograph-like image. Different approaches to model projection are presented. Despite the noise which comes from the high degree of interpolation in the model used, the generated results are the first step in the generation of panoramic craniofacial images without the use of ionizing radiation.

Simultaneous EEG-fMRI measurements are used to study epileptic patients in order to localize brain regions associated with interictal epileptiform discharges (IEDs). The common approach to this analysis uses the generalized linear model (GLM) with a canonical hemodynamic response function (HRF) to relate EEG activity to predicted BOLD responses. The study aimed to determine whether the canonical HRF is optimal for IED-induced hotspot fMRI analysis and develop optimized HRFs. The optimization of four HRF models was performed on BOLD responses derived from SPM activation clusters from 36 epileptic patients. The results showed that the optimized HRF models improved the sensitivity of the EEG-fMRI analysis of IED events in epilepsy patients compared to standard HRFs. This was evidenced by an increase in the size, number, and maximum t-score values of activated areas.

Spatial resolution, signal-to-noise ratio (SNR) and acquisition time are interconnected in magnetic resonance imaging (MRI). Trade-offs are made to keep the SNR at the acceptable level, maximizing the resolution, minimizing the acquisition time and maintaining radiologically useful images. In low-field MRI scanners and especially in diffusion imaging, these trade-offs are even more crucial due to a generally lower image quality. Image post-processing is necessary in such cases to improve image quality. In this work, we alleviate the challenges of low SNR in dMRI at low magnetic fields by performing super-resolution reconstruction (SRR). Our approach combines multiple low-resolution images acquired at different image slice rotations and employs a convolutional neural network to perform the SRR. Training is performed on noisy images. The network learns to extract and compose complementary image details into a super-resolution output image. Because of the properties of noise and the training process, the super-resolution images are less noisy than the directly acquired high-resolution ones, contain more high-resolution details than the input low-resolution images and the total acquisition time is decreased.

Radiomic features, numeric values extracted from a region of interest (ROI) in medical images, can be used to train prognostic models for various types of cancer. However, in locally advanced diseases, more than one lesion may be present. Using the information contained in multiple regions increases the complexity and necessitates additional processing. Here, we tested seven strategies of handling multiple regions in radiomic-based regularized Cox regression for predicting metastasis-free survival using a cohort of 115 non-small cell lung cancer patients. We have found that using all ROIs to fit the model allowed for better results than using only the largest ROI, achieving c-indexes of 0.617 and 0.581, respectively.

Locally advanced non-small cell lung cancer (NSCLC) is characterized by poor prognosis and is the leading cause of cancer-related deaths worldwide. The main problem in NSCLC is treatment resistance and a high potential for progression and metastasis, which results in low survival rates for patients. Immunotherapy (IO) alone or combined with chemotherapy and targeted therapies, while dedicated to selective subgroups of patients, replaced traditional cytotoxic therapies (radiotherapy or chemotherapy) in advanced stages of disease. Therefore, it’s important to search for new features that may help predict the effectiveness of currently available therapies. The paper presents preliminary results of exploratory data and survival analyses (Kaplan-Meier survival curves, uni- and multivariate Cox proportional hazards models), indicating potential factors that may help stratify the group of patients to select the appropriate treatment.

The major cause of cancer patient death is the development of distant tumors (metastases) that leads to a disease that is hard to treat. Therefore, the emergence of metastasis is a key time-point in the management of cancer leading to a decision about more intensive treatment. One of the most metastatic cancer types is lung cancer, in which the most common subtype is non-small cell lung cancer (NSCLC). We selected a set of common genetic variants in 27 genes and identified them in DNA from the peripheral blood of 335 NSCLC patients. Our goal is to apply a predictive machine learning model to estimate the risk of metastasis based on clinical and SNPs data. We have found that Cox regression with LASSO regularization gives the best predictive accuracy with a concordance index equal to 0.63.

Modelling of cancer growth with metastasis can be approached in different ways. The aim of this work is to compare two model types to answer a question of their equivalence and applicability. The first model is a compartmental one, described by ordinary differential equations, while the other is a distributed parameter model, given by partial differential equations. The comparison is based on the overall number of metastatic cells, changing in time. Various initial tumour sizes, growth rates and time horizons are taken into account.

A deterministic model consisting of both an ordinary differential equation and delay differential equations was proposed to describe the problem of lung cancer growth and its metastatic progression. In order to represent the behaviour of a virtual population of affected individuals, some of the model parameters are randomly generated. This approach makes it possible to determine certain characteristics describing the population, such as the classification of the primary cancer, location of the metastasis or the estimation of survival curves.

The act of engaging in secondary activities while driving can cause safety risks on public roads due to the driver’s distracted attention. The objective of the research was to predict changes in driver concentration levels caused by secondary activities (eating, drinking, bending, and turning toward the rear seats) using the electrooculographic (EOG) signal. Four subjects, consisting of one male and three females between the ages of 23 and 57, performed distracting driving activities using a driving simulator. The EOG signals were recorded using JINS MEME Academic Pack smart glasses, and machine learning techniques (boosted trees, bagged trees, subspace discriminant, subspace KNN, RUSBoosted Trees) were used to classify the occurrence of secondary activities. The results show that the highest accuracy (87%) has been achieved for the bagged tree (ensemble classifier).

This work aims to develop a numerical tissue model and implement software to simulate photon propagation using the Monte Carlo method to determine design guidelines for a physical measurement system. C++ was used for the simulation program, and Python as a programming environment to create an interface that allows the user to customize individual simulation elements, allowing for increased accuracy and flexibility when simulating photon movement. This allows the user to customize the simulation to their specific requirements, ensuring the results are as accurate and reliable as possible. It also models the detector to determine if a given photon is in the desired location. The program simulates the propagation of light from a normal illumination medium with anisotropic scattering and records the escape of photons on the upper surface. The simulation also takes into account absorption and scattering coefficients for a given wavelength, and data regarding these parameters are read from a .csv file. The variance reduction technique is used to improve the efficiency of the simulation. The user interface allows users to define their own parameters, such as wavelength, anisotropy coefficient, refractive index, and layer thickness. In this paper, we simulate four photodiodes and different distances between the source and detector to determine the most suitable model for designing a physical sensor.

This article presents methods for surface reconstruction and volume determination based on the point cloud created by 3D imaging. Such a system would be used to accurately estimate breast volume in patients classified for breast reconstruction surgery at plastic surgery centers. To develop such a system, various methods of determining volume, based on images from the Intel D435i camera, were tested. In addition, an application and a measuring station tailored to clinical needs were developed. Finally, we stated that 3D imaging systems can effectively determine breast volume for surgical procedures.

This paper describes a research study that investigates the use of machine learning algorithms on synthetic data to classify the risk of developing preeclampsia by pregnant women. Synthetic datasets were generated based on parameter distributions from three real patient studies. Four models were compared: XGBoost, Support Vector Machine (SVM), Random Forest, and Explainable Boosting Machines (EBM). The study found that the XGBoost and EBM consistently outperform the other models. An analysis of patient subsets based on their pregnancy history was also conducted, revealing that the group of patients in their first pregnancy achieved the highest prediction accuracy. Additionally, the study explored the efficacy of risk prediction based on various parameters and found that the results vary depending on the models used and the degree of class balance in the database. Finally, an additional test was performed on the dataset annotated by physicians.

The purpose of this research was to prepare a preliminary mathematical approach for the identification and analysis of the gap distribution using changes in brightness on CT and spatial analysis of the aorta. We used AngioCT data from three male patients, 45 to 65 years of age, with acute type IIIb aortic dissection that originated near the left subclavian artery and involved the renal arteries. The analysis was performed on CT slides from the descending aorta to the diaphragm. Each time Feret DF diameter, width of gap in dissection and average brightness BAV were analyzed. Finally, to describe the influence of the gap distribution in the area of aortic wall dissection, a mathematical function was calculated that combined the difference in brightness value and the gap width for each CT scan. The results showed that there is a positive correlation between the width of the gap (tear) within the dissected aorta and the brightness coefficient. The gathered results also stress that the most accurate data are when the gaps in the dissected wall are equally distributed throughout the dissection and are not concentrated in a few areas within the dissection. In summary, our data show that there may be a correlation between the width of the gap and the brightness coefficient calculated from CT scans that may also be used for the prognosis of hemodynamic changes within the dissection. In addition, we also highlight the need for the appropriate number of patients incorporated for computational analysis of spatial configuration of gaps in the dissected aorta, as too few cases can lead to misleading results.

Recent advancements in artificial intelligence have brought some spectacular innovations, including systems for automatic analysis of respiratory sounds. These solutions, however, usually have a very limited set of supported signal sources and it raises question if these can be adapted to work with signals that they have not been trained on. This work explores the possibility to automatically discover the characteristics of unknown source of body signals and convert them to match that of a known source. Next, it answers a question of how well the AI can perform given signals processed in this way. Our proposed method is used to adapt the ICBHI 2017 Respiratory Sounds Challenge recordings to match the expected input signal characteristic of commercially available AI solution for lung sounds analysis. The method achieves an ICBHI score of 60.99% which is an improvement of 2.7% points over the current challenge leader. The presented results show that existing AI solutions can solve congenial tasks even when fed with signals from unknown sources and even achieve superior performance to dedicated solutions. In this case the advantage of using larger and more precisely described training data during AI development outweighs potential imperfections of the audio data transformation leading to higher accuracy. In fact, the true performance of the proposed solution might be even higher because some ICBHI recordings were found to be mislabeled, which could also explain limited success of models developed based on this data.

Bioelectrical impedance analysis (BIA) consists in measuring the total resultant electrical resistance of the body, which is a derivative of the resistance and reactance, by using a set of surface electrodes, connected to a computer analyzer and using a current of a given frequency and intensity. Bioimpedance spectroscopy (BIS) is an extension of BIA method that involves measuring the impedance of the test object over a range of frequencies. The aim of this systematic review was to evaluate prevalence of bioelectrical spectroscopy in medicine, especially (a) for measuring blood glucose level (b) classification of pulmonary nodules (c) detection and estimation of lymphedema and breast cancer-related lymphedema (BCRL). We focus on finding a new area of activity, where we can use BIS and BIA, which is a critical aspect in the context of rapidly developing medicines. A systematic research on electronic databases (PubMed, Cochrane) from inception to January 2018 was performed according using PICO strategy. The quality of the publication (clinical trials) was evaluated by using Jadad checklist. Through the initial literature search and after removing duplicates and excluding papers by screening titles and abstracts, 140 potentially relevant studies were examined. Fourteen studies met the inclusion criteria. The review confirmed that BIA and BIS are used as a tool to assist in the diagnosis of patients to a very limited extent. In addition, it is worth mentioning the lack of clinical reviews based on the application of machine learning or deep learning to narrow pathologies.

Diagnosing neurological disorders and tracking rehabilitation progress requires objective, quantitative measures. Isokinetic exercises are commonly employed to evaluate muscle strength and activity. However, repetitive, quantitative assessment of the tests performed during these exercises can be difficult, especially for patients with neurological disorders. Usually, either dynamometric or sEMG measurements are used. This study is concerned with the evaluation of the applicability and reliability of these two measurement techniques, based on analysis of elbow flexion and extension supported by the Luna EMG rehabilitation robot, aimed at the evaluation of the activity of the biceps and triceps muscles. Dynamometric and surface electromyography measurements are compared with respect to their repeatability between sessions and days, and their ability to differentiate impaired limbs from healthy ones.

Rehabilitation of patients with impaired grip functions involves standard exercise that should be repeated in a predefined sequence. They are aimed to gradually improve patients’ ability to grip and operate objects used in daily life. Rehabilitation robots, employing surface EMG (sEMG) as a diagnostic and feedback signal may be used to support these exercises and trace the rehabilitation progress. However, to properly employ exercise regimen, information about what force should be used to hold and operate various objects is needed. This work is focused on finding the correlation between the value of sEMG signal and the force required to grip a particular object. Calibration plots obtained that way will reduce the need for additional equipment, such as specialized gloves, and at the same time provide information needed for better exercise protocols.

This article presents the results of a study on the classification of 36 letters of the Polish Sign Alphabet (PSL) using a data glove. The data glove includes a Raspberry Pi, a PCB with four 16-bit ADS1115 ADCs, and ten piezoresistive sensors. The sensors function as potentiometers that respond to finger bending. A neural network consisting of LSTM and convolutional layers was used for classification, achieving an efficiency of 99% on data from a single subject, which underwent prior augmentation. The study suggests that the device could potentially be calibrated solely from the target user. This is the first study of its kind on PSL.

This paper analyzes the data concerning the elderly and the aspects of their lives that include taking medications and then proposes a device that was created in order to help elderly organize their medications. A solution consisting of mobile application and a 3D-printed box with Arduino board is proposed. Mobile application is responsible for gathering and processing all the information received from the Arduino, like status of the box (opened or closed), temperature, humidity, and pressure. It is also responsible for saving all new user’s medications and displaying them in a form of calendar and list of medications for given day. When the box gets opened, the right medication gets automatically set as taken. The tests made by users showed that the whole product was received really well. Most suggestions concern the software of the product, and they can be easily implemented with software update.

Telemonitoring of vital signs combined with other multimodal health data can provide a basic foundation for data medicine. The collected data can then be analyzed semi-automatically and made available to the physician for better assessment of the state of health. This procedure was tested in use-cases in a nursing home and with elderly people at home. In addition to several vital signs, patients were able to submit details of their health status. To enable a high number of valid measurements, support measures were established that are adapted to residents in nursing homes, in the home, and caregivers.

Continuous monitoring of cardiac function is highly desirable for long-term assessment of cardiovascular health, detection of acute cardiac dysfunction. Information from volume flow rate may be important in a variety of different clinical circumstances, such as stroke, arteriovenous malformation, cardiac failure. The main aim of the work was review of non-invasive measurements of carotid blood flow methods and devices. There are various types of methods and devices for monitoring blood flow in the carotid artery, however, so far no mobile recorder of the appropriate quality of registration has been developed. The most promising solutions are wireless wearable patches that, when applied to the skin of the neck, will not be a burden during measurements. This is important when measurements are to be made, during extreme activities, such as in pilots. As a result of the conducted research, there is a presumption that carotid blood flow measurements methods has great potential to be a useful tool in the diagnosis of the pilot's predisposition, at the same time performing a warning function, the so-called alerters, before the imminent overload loss of consciousness.

This study explores the feasibility of unsuppressed water (WU) peaks acquired as a part of prescanning in conjunction with metabolites from suppressed water (WS) spectra for temperature measurements. Calibrations and in vivo single-voxel 1H MRS was conducted on a 3-T GE scanner. Calibration constants for the water-chemical shift were obtained by using a temperature-controlled phantom containing an aqueous solution of N-acetyl aspartate (NAA), Creatine (Cr), and Choline (Cho). Estimations of absolute human brain temperature were performed utilizing the correlation of temperature to the water-chemical shift for the NAA, Cr, and Cho resonances. Commercially available single-voxel point-resolved spectroscopy (PRESS) sequences (repetition/echo time = 1500 ms/30 ms; voxel size 2 × 2 × 2 cm3) were used to acquire data for calibration of the metabolites’ chemical shift differences and in vivo temperature estimations. Each sequence included 16 averages without water suppression and 96 averages with water partially suppressed. In vivo study consisted of 2 PRESS sequences, one before and one after extensive 30-min fMRI task acquisition. The mean brain temperatures calculated using NAA, Cr, and Cho with WS were 37.49 ∓ 0.40 ℃, 37.09 ∓ 0.34 ℃, and 38.18 ∓ 0.37 ℃ before fMRI and 37.19 ∓ 0.45 ℃, 36.81 ∓ 0.39 ℃, and 37.88 ∓ 0.41 ℃ after fMRI, with a difference of approximately 0.3℃. When using WU with NAA, Cr, and Cho, the mean brain temperature was 38.13 ∓ 0.49 ℃, 37.67 ∓ 0.44℃, and 38.86 ∓ 0.52 ℃ before fMRI and 37.89 ∓ 0.52 ℃, 37.45 ∓ 0.50 ℃, and 38.62 ∓ 0.52℃ after fMRI, with a difference of approximately 0.24℃. Similar differences reported by using both methods for temperature calculation suggest that using unsuppressed water for temperature measurement in absence of its suppressed counterpart is feasible and may add another angle to already concluded past MRS studies which is temperature.

The problem of navigating multi-robot systems through environments in a collision-free manner is becoming ever more important in the healthcare industry. An increasing interesting can be observed in optimization-based methods for path planning in multi robot systems. This study analyzes the performance of a Real-Coded Genetic Algorithm when planning paths for multi-robot systems with up to five robots. Our approach uses the control location method extended to multiple robot systems. The entire path-planning task was contained within the objective function, which meant that no custom modifications to the optimization algorithm were required, furthermore no constraints were placed on the robot displacements—the collision space was continuous. In total 625 simulations were performed over five custom collision maps and including five reruns of the procedure to minimize the effects of random number generator. Our findings indicated that the methods was viable in cases up to five robots, where the average success rate for path planning was more than 37% with a low number of generations and population size. Depending on the map the success rate could be as high as 80% for systems with 4 robots. Interestingly, avoiding robot-to-environment collisions was more difficult than robot-to-robot collisions, while the number of control locations proved to be more costly for the algorithm than the number of the robots. This might suggest that the algorithm is more suited to simpler collision spaces with higher number of robots.

Every year in many countries, there are fatal unintentional drownings in different water reservoirs like swimming pools, lakes, seas, or oceans. The existing threats of this type require creating a method that could automatically supervise such places to increase the safety of bathers. This work aimed to create methods and prototype solutions for detecting people bathing in water reservoirs using a multimodal imaging system and machine learning. Two types of cameras, RGB and thermal, were integrated and calibrated to form a multimodal imaging system. The system was designed and implemented to acquire real-world data for bathing people in swimming pools. The EfficientDet models were adapted and trained on collected data reaching at least 94% detection accuracy, with the highest result equal to 97.47%. The best accuracy obtained for the thermal data was lower and equal to 94.85%. However, thermal imaging allows observing scenes in low-light conditions or darkness. This could potentially highly improve the effectiveness of rescue missions, decreasing the death rates or improving the health of early rescued people. Thermal imaging could also be more acceptable regarding privacy, as high-frequency biometric features are not as easy to extract from thermal images as from high-resolution RGB images.

This paper reports on an assistive system for the visually impaired that presents the environment through haptic stimulation. This wearable system consists of a processing unit attached to the user's chest and a haptic belt placed to the user’s abdomen. The custom-made belt consists of 20 vibration actuators arranged in a 4 × 5 matrix. The depth images captured by the camera are processed to identify the orientation of the ground plane and the location of obstacles. The system's detection range is limited to 2.5 m within a user-defined corridor 0.8−1.5 m wide. The segmented obstacles in the depth images are projected onto the so-called occupancy grid, which is a top view of the scene. Each grid cell corresponds to a haptic actuator that is activated when the cell is occupied by an obstacle. The system is capable of real-time haptic representation of the environment. We present successful indoor technical trials of the system, demonstrating its validity as a potentially useful travel aid for the visually impaired.

Long-term exposure of polymer dental composites to selected liquids of different chemical composition and characteristics (pH level, sugar content, etc.), like: coffee, red wine, soy sauce, alkaline mineral water, coke drinks and mouthwash liquid, was studied. The degree of chemical and physical changes to the surface layer of the materials were evaluated based on the results determined by infrared spectroscopy, colorimetry, microhardness, microroughness and wettability. The experiments confirmed suspicions about the liquids studied adversely affect polymer dental composites. It was expected that the material is generally not resistant to the aging agents present in beverages. However, quite unexpectedly, similar behavior was also exhibited by the liquid for oral hygiene. Diet coke, alkaline water and mouthwash liquid had the biggest impact on wear of the composites, contrary to soy sauce. Colorimetry proved the greatest color changes for the samples treated with red wine and coffee, whereas their incubation in alkaline mineral water and mouthwash liquid, made them clearly brighter. Chemical modification of the their surface layer, influenced their microhardness and, together with changes to the surface geometry, wettability. Intriguing differences were observed between regular coke and diet one. Coke with sugar penetrated the composite samples to a smaller depth in comparison to its sugar free version. The observations are important in the context of the selection of polymer composite materials for a specific dental case and the awareness of patients regarding discoloration of teeth and composite fillings, under the influence of liquids present in the daily diet and hygiene.