Developmental hip dysplasia (DDH) is a disease in which the hip joint fails to develop normally due to various causes before, during or after birth. The most important method used for the detection of DDH is hip ultrasonography. The stage of obtaining the hip US image varies because it depends on the operator and external influences. In this study, an artificial intelligence-based system has been developed to eliminate this variability. The developed system includes a 2-stage deep learning model. The main purpose of the system is to automatically determine whether the US images obtained by physicians are suitable for the calculation of alpha and beta angles required for diagnosis. The system uses the U-NET architecture in the first stage and the masked region-based convolutional neural network (MBT-ESA) architecture in the second stage. For the training, 540 images were taken from Sel & ccedil;uk University Faculty of Medicine hospital with the approval of the ethics committee. A total of 840images were obtained for training with data augmentation. U-NET architecture training resulted in an accuracy of0.93 and region-based convolutional neural network training with mask resulted in an accuracy of 0.96. The overall system accuracy was calculated as 0.96. The results obtained in this study suggest that by increasing the number of real-time tests and images, the inter-operator variability in the diagnosis of DDH can be eliminated. Translatedwith DeepL.com (free version)

Background and Objective: Neonatal health is critical for early infant care, where accurate and timely diagnoses are essential for effective intervention. Traditional methods such as physical exams and laboratory tests may lack the precision required for early detection. Hyperspectral imaging (HSI) provides non-invasive, detailed analysis across multiple wavelengths, making it a promising tool for neonatal diagnostics. This study introduces HarmonyNet, an involution-based HSI model designed to improve the accuracy and efficiency of classifying neonatal health conditions. Methods: Data from 220 neonates were collected at the Neonatal Intensive Care Unit of Sel & ccedil;uk University, comprising 110 healthy infants and 110 diagnosed with conditions such as respiratory distress syndrome (RDS), pneumothorax (PTX), and coarctation of the aorta (AORT). The HarmonyNet model incorporates involution kernels and residual blocks to enhance feature extraction. The model's performance was evaluated using metrics such as overall accuracy, precision, recall, and area under the curve (AUC). Ablation studies were conducted to optimize hyperparameters and network architecture. Results: HarmonyNet achieved an AUC of 98.99%, with overall accuracy, precision and recall rates of 90.91%, outperforming existing convolution-based models. Its low parameter count and computational efficiency proved particularly advantageous in low-data scenarios. Ablation studies further demonstrated the importance of involution layers and residual blocks in improving classification accuracy. Conclusions: HarmonyNet represents a significant advancement in neonatal diagnostics, offering high accuracy with computational efficiency. Its non-invasive nature can contribute to improved health outcomes and more efficient medical interventions. Future research should focus on expanding the dataset and exploring the model's potential in multi-class classification tasks.

Breast cancer screening demands accurate, non-invasive, low-cost tools. Infrared thermography is radiation-free and portable, but its utility hinges on robust computer-aided diagnosis (CAD). We benchmark three deep-learning families for static multi-view breast thermography—CNNs, Transformers, and an involution-based model (HarmonyNet-Lite). Experiments use the Breast Thermography dataset (119 patients; 476 manually segmented ROIs from anterior/oblique views). A compact pipeline performs ROI segmentation, RGB conversion, normalization, resizing, and moderate data augmentation; class imbalance is handled with minority oversampling and class-weighted loss. Evaluation follows patient-stratified five-fold cross-validation. HarmonyNet-Lite yields the best results: accuracy 87.47 ± 2.99

Ponticulus posticus (PP) is a bony structure in the cervical spine, often difficult to identify in radiographic images, and its detection is important for both orthodontic diagnosis and clinical decision-making related to craniovertebral pathologies. The purpose of this study is to develop a deep learning-based approach for detecting the PP in lateral cephalometric radiographs using the YOLOv8-seg model. This retrospective study analyzed a dataset of 1000 anonymized lateral cephalometric radiographs, focusing on the segmentation and detection of the PP. Images were resized to 640 × 640 pixels and labeled by two experienced dentomaxillofacial radiologists. The YOLOv8-seg model, designed for segmentation tasks, was trained over 100 epochs with a batch size of sixteen, using pre-trained weights from the COCO dataset. Model performance was evaluated using precision, recall, mean average precision (mAP), and F1 score metrics. The YOLOv8s-seg model demonstrated high accuracy in detecting the PP, with a precision of 62.81

Early detection and accurate diagnosis of neonatal diseases are crucial for improving health outcomes and reducing infant mortality. This study introduces a novel Hybrid Convolutional and Involutional Spectral Network (HybridCISN) that integrates hyperspectral imaging (HSI) data with blood biomarker analysis to enhance neonatal health diagnostics. By combining 2D convolution, 3D convolution, and involution layers, the HybridCISN model extracts spatial, spectral, and channel-specific features, addressing limitations in traditional diagnostic methods. The model was evaluated through two distinct approaches: (1) using only HSI spectral data and (2) integrating HSI spectral data with blood biomarkers such as haemoglobin and bilirubin levels. These approaches were tested for both binary classification (healthy vs. unhealthy neonates) and multiclass classification (specific neonatal diseases such as intracranial hemorrhage, necrotizing enterocolitis, pneumothorax, and respiratory distress syndrome). Experimental results demonstrate the HybridCISN model's superior performance, achieving an overall accuracy of 93.64% for binary classification and 90.25% for multiclass classification. Compared to state-of-the-art methods such as the involution-based HarmonyNet and the 2D/3D convolution-based HybridSN, the HybridCISN model achieved accuracy improvements of 0.8% and 1.5%, respectively, in multiclass classification. The second approach, integrating blood biomarkers, improved diagnostic sensitivity and specificity, emphasizing the value of multimodal data fusion. Involution layers reduced channel redundancy and optimized feature extraction, as confirmed by ablation studies. The HybridCISN model offers a scalable and non-invasive diagnostic framework, addressing clinical applicability and biomarker accessibility, while combining precision, efficiency, and robustness to advance neonatal disease detection and set a benchmark for future research in medical imaging.

Early diagnosis of lower extremity injuries in professional football players is crucial for maintaining performance and minimising long-term risks. Despite the growing use of thermographic imaging as a non-invasive tool for detecting musculoskeletal disorders, its integration into automated injury detection systems remains limited, particularly under data-scarce conditions. Given the need for effective early detection methods and the potential of thermography in sports medicine, this study investigates the applicability of deep learning models for classifying lower extremity injuries. Specifically, it evaluates the performance of Prototypical Network and Siamese Network models using thermographic data collected from professional athletes. The original dataset consists of images from 16 healthy and 9 injured individuals, and through augmentation it was expanded to 360 healthy and 180 injured samples. The Prototypical Network achieved an accuracy of 97.78%, while the Siamese Network attained 94%. These findings indicate that both models are capable of accurate injury detection, despite challenges posed by class imbalance and limited data availability. In conclusion, the study highlights the effectiveness of thermographic imaging combined with deep metric learning in identifying injuries in professional football players and suggests that reliable results can be achieved even in constrained data environments.

This study presents a comprehensive evaluation of deep learning methods for classifying strokes using computed tomography (CT) images. Focusing on the comparative analysis of three ResNet architectures—ResNet-50, ResNet-101, and ResNet-152—the models were trained on a specialized dataset provided by the Ministry of Health of the Republic of Turkey. This dataset comprises meticulously annotated brain CT scans, validated by a team of experienced radiologists to ensure precision and reliability in identifying stroke regions. The performance of each model was assessed using several key metrics, including accuracy, precision, recall, and F1 score, to provide a holistic evaluation. Among the models, ResNet-152 demonstrated the highest efficacy, achieving an accuracy of 97.24

The Objective: This study aimed to evaluate, at two time points (the 1st and 14th training days), the effects of training provocation on thermal asymmetry in football players with and without a lateral ankle sprain (LAS) injury history. Methods: Twenty-seven football players from the U-19 squad of a Turkish Süper Lig club were included. Athletes were divided into two groups by injury history: with LAS injury history (n = 10) and without LAS injury history (n = 17). On the 1st and 14th training days, pre-training and post-training infrared thermographic images were obtained. For the ankle, patellar tendon, calf medialis, calf lateralis, and tibialis anterior regions, the side-to-side temperature difference (ΔT) and the post-training change (ΔPost–Pre) were calculated. Results: On the first day, athletes with LAS injury history showed a marked increase in ΔT at the ankle (+0.19 ℃) and patellar tendon (+0.22 ℃), whereas a decrease was observed in the control group. On the fourteenth day, ΔT values were elevated from pre-training in the injury-history group and expanded toward the calf muscles (calf medialis +0.07 ℃; calf lateralis +0.06 ℃). Tibialis anterior exhibited a decrease in both groups. Conclusion: In football players with LAS injury history, thermal asymmetry that emerges acutely at the joint–tendon level with training load extends to the muscle level over two weeks. This pattern indicates lasting alterations in neuromuscular control and load distribution. AI-assisted thermography can sensitively reveal such asymmetries and may serve as a valuable tool for individualized load management and injury prevention strategies during return-to-play.

The RGB color ring is known as the most understandable color representation in human vision, as it has complementary colors. However, color relationships hardly ever play a function in waveletprimarily based totally color image processing tools. In this study, Complementary Color Wavelet Transform (CCWT), which is supported by complementary color relationships and complex wavelet design techniques, is used to denoise in color images. This wavelet consists of a family of two-dimensional complex wavelets with a phase difference of 2 pi/3 obtained from the angle relationship between the color axes of the RGB color ring, and is very effective in terms of directional selectivity. By using the coefficients of the directions in different phases, denoising processes are performed from the multi -channel color images. It was validated the performance of CCWT using various color images and noise levels, based on peak signal-to-noise ratio, structural similarity index, mean square error values, and visual quality. CCWT was compared with state-of-the-art multi -resolution image denoising algorithms, and found that the method achieves superior denoising performance both quantitatively and visually. It was also analyzed the computation time of CCWT and compared it with existing approaches.

This study examines the effects of balance and strength on the thermal heat map. Participants' thermal temperature values, isokinetic values and balance values were taken in a specific measurement procedure and the relationship between them was statistically analyzed. All processes were carried out under the supervision of a physiotherapist. According to the statistical results we obtained, a significant relationship was found between isometric strength and thermal temperature values in hamstring group muscles. The relationship between isometric force and thermal temperature map of the hamstring group muscles is important in terms of showing that the hamstring muscle group should be prioritized in the analysis of thermal evaluations of basketball athletes. In addition, a significant relationship was found between balance parameters and temperature values in vastus lateralis and ankle. The thermal finding we observed in the vastus lateralis muscle and ankle region in terms of balance parameters, in addition to the evaluation of ankle injury risks of basketball athletes, has raised suspicion that the frequency of knee injuries is adaptively loaded by the vastus lateralis muscle and especially knee injuries occur as a result of this adaptation. From this point of view, it provides evidence that thermal analysis is a useful tool in the analysis of injury mechanisms and injury risks in understanding the most common ankle and knee injuries in basketball.

When a classification process is performed using Class Activation Maps, which is one of the Explainable Artificial Intelligence approaches, the areas influencing the classification on the input image can be revealed. In other words, it is demonstrated which part of the image the classifier model looks at to make a decision. In this study, a 200-class classification model was trained using the open-source dataset CUB 200 2011, and the classification results were visualized using the EigenCAM and HayCAM methods. When comparing object detection performances based on the areas influencing classification, the EigenCAM method reaches an IoU (Intersection over Union) value of 30.88%, while the HayCAM method reaches a value of 41.95%. The obtained results indicate that outputs derived using Principal Component Analysis (HayCAM) are better than those obtained using Singular Value Decomposition (EigenCAM).

Football clubs use various methods such as thermal imaging which is a non-invasive and faster method to detect injuries and increase the success rate of the football club by reducing the injury rate. Studies have proven that with thermal imaging it is possible to detect inflammation caused by an injury. Therefore, it is possible to detect potential injury with infrared thermography. One of the biggest handicaps of injury detection with thermal imaging is that it is open to subjective interpretation, there are many points that can be missed, and it takes time to analyse them one by one. In order to avoid this problem, to increase the success of injury detection, a deep learning supported pipeline has been designed in this study to detect injuries from thermal images. In this pipeline, the hamstring muscle region from the football player thermal images was segmented using U-Net architecture. After that in order to detect injuries, segmented muscle region is classified by using Densenet, Resnet, VGG, Efficientnet architectures variations and feature pyramid added at the end of these architectures. Among the architectures used for classification, the EfficientnetB0 and EfficientnetB1+feature pyramid architectures are the most successful, with accuracies of 83.9% and 81%, respectively.

Thermal imaging systems are harmless to human health and enable contactless heat measurements. The thermal cameras are used in many public sectors where it is necessary to detect the change of temperature values. However, thermal cameras are costly and produce images with low edge information. This situation prevents the widespread use of thermal cameras. Therefore, in recent years, the researches to advance the quality of thermal images have increased. Within the scope of the studies in this paper, first of all, three different datasets consisting of thermal images in the colourful format of neonates were created. Also, TSRGAN+ deep network model was presented for super-resolution studies. The super-resolution images obtained visually approached ground truth images to a great extent. In addition, these results were compared using the peak signal to noise ratio (PSNR) and the structural similarity index measure (SSIM) image quality metrics. The proposed model showed a superior success in terms of the values of PSNR and SSIM compared to the state-of-the-art models. Here, the PSNR value of the proposed TSRGAN+ model increased by 1-1.5 dB compared to the TSRGAN network architecture, while the SSIM value increased by around 2-3%. Finally, the unhealthy-healthy image classification applications were performed on all thermal image sets in order to implement both the task-based evaluation and a real-life application. Thus, a new method is presented to evaluate the results of the super-resolution studies. Here, firstly, a CNN-based classifier was designed and the classification metrics were obtained for all three datasets. Then, transfer learning was applied using state-of-the-art models (ResNet101, Xception) to increase classification success. Here, the most successful results were obtained in applications using the ResNet101 model. Also, the developed model outperformed the TSRGAN, which achieved the second-best result. When all the obtained results are evaluated, it has been observed that the super-resolution models increase the success of unhealthy-healthy classification by about 10% compared to the low resolution images. In other words, the effects of super-resolution techniques on classification applications are clearly seen. In summary, the logical use of the super-resolution research will enable the common use of low-cost thermal cameras in the application fields such as medicine.

Visual XAI methods enable experts to reveal importance maps highlighting intended classes over input images. This research paper presents a novel approach to visual explainable artificial intelligence (XAI) for object detection in deep learning models. The study investigates the effectiveness of activation maps generated by five different methods, namely GradCAM, GradCAM++, EigenCAM, HayCAM, and a newly proposed method called "HayCAMJ", in detecting objects within images. The experiments were conducted on two datasets (Pascal VOC 2007 and Pascal VOC 2012) and three models (ResNet18, ResNet34, and MobileNet). Zero padding was applied to resize and center the objects due to the large objects in the images. The results show that HayCAMJ performs better than other XAI techniques in detecting small objects. This finding suggests that HayCAMJ has the potential to become a promising new approach for object detection in deep classification models.

Deep learning models are proficient at predicting target classes, but they need to explain their predictions. Explainable Artificial Intelligence (XAI) offers a promising solution by providing both transparency and object detection capabilities to classification models. Mask detection plays a crucial role in ensuring the safety and well-being of individuals by preventing the spread of infectious diseases. A new visual XAI method called HayCAM+ is proposed to address the limitations of the previous method known as HayCAM, such as the need to select the number of filters as a hyper -parameter and the use of fully -connected layers. When object detection is performed using activation maps created via various methods, including GradCAM, EigenCAM, GradCAM++, LayerCAM, HayCAM, and HayCAM+, it is found that HayCAM+ provides the best results with an IoU score of 0.3740 (GradCAM: 0.1922, GradCAM++: 0.2472, EigenCAM: 0.3386, LayerCAM: 0.2476, HayCAM: 0.3487) and a Dice score of 0.5376 (GradCAM: 0.3153, GradCAM++: 0.3923, EigenCAM: 0.5003, LayerCAM: 0.3928, HayCAM: 0.5098). By using dynamical dimension reduction to eliminate unrelated filters in the last convolutional layer, HayCAM+ generates more focused activation maps. The results demonstrate that HayCAM+ is an advanced activation map method for explaining decisions and detecting objects using deep classification models.

Manual segmentation of patient CT images is both time-consuming and labor-intensive. Additionally, classic image processing techniques are insufficient in CT images due to the close pixel values of tissues. Automatic segmentation of the aorta in human anatomy can reduce healthcare workers' workload in preoperative planning. This study compares the performance of the AKG-Unet segmentation model with other models (U-Net, Inception UNetv2, LinkNet, SegNet, and Res-Unet) on thoracic aorta, abdominal aorta, and iliac arteries segmentation in contrast CT images. Initially, pixel intensities in the Kits and Rider datasets were recalibrated. Then, 2D axial images underwent resizing and grayscale normalization. Segmentation models have been trained and tested with 5-fold cross-validation. 2D prediction masks were stacked to generate a 3D output, and spatial information was transferred to the predicted mask. In the 3B aortic segmentation, small objects adjacent to it were removed using image processing techniques. In our study, the AKG-UNET model achieved the highest segmentation results on the AVT dataset with a Dice score of 91.2%, Intersection-Over-Union (IoU) score of 85.6%, sensitivity of 90.9%, and specificity of 99%. A method has been proposed that helps physicians analyze the aortic structure, and segments the aortic structure so that they can intervene in the correct location and make a preoperative evaluation.

Convolutional methods are commonly used for hyperspectral imaging (HSI) classification. However, HSI datasets are large due to numerous narrow-band spectra, leading to high computational costs and optimization challenges in convolution-based deep learning models. To address this, we propose the involutional residual spectral network (IRSN), using involution kernels tailored to the data for meaningful feature extraction. IRSN achieves this with fewer parameters than convolutions. By leveraging involution layers based on spectral signatures, IRSN captures spectral-spatial information. Furthermore, residual blocks within the network facilitate information preservation and overcome gradient-related challenges. Experimental studies conducted using four publicly available datasets demonstrate that the proposed IRSN model outperforms certain state-of-the-art convolutional-based networks in terms of effectiveness and efficiency.

Early diagnosis systems have vital importance to monitor and follow-up the conditions of neonates. Thermal imaging as a non-invasive and non-contact method has been used to monitor neonates for over decades. In this study, we train a convolutional neural network (CNN) model that classifies medical thermograms as healthy and unhealthy using real neonatal thermal images captured within a year from the Neonatal Intensive Care Unit (NICU), Faculty of Medicine at Selcuk University, Turkey. The trained model achieved 99.91% accuracy for train, 99.47% accuracy for validation, and 99.82% accuracy for test data. The test data were never used during training. Although the trained model achieves over 99% accuracy, how it works was not known because of the CNNs' "Black-Box" nature. The four visual Explainable Artificial Intelligence methods that are GradCAM, GradCAM++, LayerCAM, and EigenCAM and a new ensemble visual explanation method named CodCAM are used to visualise the important parts of the neonatal thermal images for classification. Therefore, medical specialists are going to know which regions of the thermograms (i.e. parts of the neonates) affect the trained CNN's decision so as to build trust in AI models and evaluate the results.

Segmentation is the process of distinguishing the desired area in an image from the background and other objects. With the development of deep learning methods, the importance of segmentation has increased, and it is now used in many fields such as medicine, industry, and autonomous systems. In this study, binary segmentation was performed on a dataset prepared with human lower extremity thermal images, and the detection of specified regions was achieved. Five different deep learning-based models were specifically designed for the problem and trained using the cross-validation method. The obtained results were recorded, and their performances were compared. Among the created models, the MCRNet model achieved the best result on the test data with a 97

A computed tomography (CT) scan is an important radiological imaging method in diagnosing pneumonia caused by SARS-CoV-2. Within the scope of the study, three classes of automatic classification - COVID-19 pneumonia, healthy, and other pneumonia - were carried out. Using deep learning as a classifier, a total of 6,377 CT images were used, including 3,364 COVID-19 pneumonia, 1,766 healthy, and 1,247 other pneumonia images. A total of seven architectures, including the most recent convolutional neural network (CNN) architectures, MobileNetV2, ResNet-101, Xception, Inceptionv3, GoogLeNet, EfficientNetB0, and DenseNet201, were used in the study. The classification results were obtained using the CT images, and they were calculated using the feature images obtained by applying local binary patterns on the CT images. The results were then combined with the help of a pipeline algorithm. The results revealed that the best overall accuracy result obtained by using CNN architectures could be improved by 4.87% with a two-step pipeline algorithm. In addition, significant improvements were achieved in all other measurement parameters within the scope of the study. At the end of the study, the highest sensitivity, specificity, accuracy, F-1 score, and Area under the Receiver Operating Characteristic Curve (AUC) values obtained for the COVID-19 pneumonia class were 0.9004, 0.8901, 0.8956, 0.9010, and 0.9600, respectively. The highest overall accuracy value was 0.8332. The most important output of the work carried out is the demonstration that the results obtained with the most successful CNN architectures used in previous studies can be significantly improved thanks to pipeline algorithms.