With ongoing advancements in Machine Learning (ML) algorithms like Convolutional Neural Networks (CNNs), their effectiveness in medical image analysis, particularly for detecting lung diseases in Computed Tomography (CT) scans, continues to improve. Although early lung nodule detection is vital, many state-of-the-art (SOTA) methods do not support simultaneous identification and classification. This separation can lead to inefficiencies and delays in clinical workflows, especially when real-time diagnosis is required. A unified approach that integrates both tasks is therefore highly desirable. Hierarchical classification models can address this by handling multi-step tasks and utilizing hierarchical label structures. This work proposes a hierarchical image classification approach for lung CT images, capable of distinguishing between nodule and non-nodule cases in the first stage and further classifying detected nodules into adenocarcinoma, small cell carcinoma, large cell carcinoma, and squamous cell carcinoma cases in the second stage. Key contributions include a method for the creation of a custom hierarchical dataset by merging the LIDC-IDRI and the Lung-PET-CT-Dx and the development of a two-level hierarchical classification algorithm with independently optimized subclassifiers. The results of this work demonstrate the effectiveness of the hierarchical model. When evaluating for simultaneous nodule identification and classification, the proposed approach achieves a weighted accuracy of 96.76%, surpassing the performance of separately executed steps. Furthermore, compared to a fat classification model, which achieves a weighted accuracy of 94.40%, the hierarchical approach outperforms it by 2.30 percentage points, highlighting the advantages of utilizing hierarchical label structures.

Lung cancer is a leading cause of death, where early detection is one of the only chances of survival. Manual early detection is time-consuming and error-prone, as it relies on the ability of radiologists to detect small nodules on hundreds of Computed Tomography (CT) images. To address the limitations of human analysis, Computer-Aided Detection (CAD) systems based on Convolutional Neural Networks (CNNs) can assist radiologists but require large datasets, struggle with image variations and lack interpretability. A Capsule Network (CapsNet) can offer an alternative, preserving spatial hierarchies and needing fewer training samples. This work presents a three-stage hybrid model combining You Only Look Once (YOLO) and a CapsNet for lung nodule detection and classification. The pipeline includes YOLOv11 for initial detection, a CapsNet for verification, and YOLOv11 for improved detection. The approach is evaluated on a combined LIDC-IDRI and Lung-PET-CT-Dx dataset. Experiments were done to compare the performance of a Full Image CapsNet classifier versus a Cropped Image CapsNet classifier, where only the Region of Interest (ROI) from detected nodules was used for classification. Results indicate that the Cropped Image CapsNet achieves superior performance with a classification accuracy of 95.57%, compared to 93.07% for the Full Image CapsNet. Despite improved precision, the three-stage pipeline shows slightly lower overall detection accuracy (89.98%) than standalone YOLOv11 (90.79%) due to increased False Negatives (FNs). While a CapsNet enhances classification reliability for FNs, further improvements in bounding box generation and a CapsNet architecture are needed to mitigate its higher False Positive (FP) rate. Future research should optimize the CapsNet, explore alternative detection models, and apply multi-class classification to distinguish benign from malignant nodules. The findings of this study contribute to the ongoing efforts in improving automated lung cancer diagnosis, offering a approach that leverages both CNN-based object detection and a CapsNet’s spatial feature encoding capabilities to enhance nodule detection accuracy and reduce radiologist workload.

The rapid evolution of football (soccer) analytics has been driven by advances in structured data analysis and recently also by Large Language Models (LLMs). However, existing methods often fail to adapt dynamically to evolving queries and lack contextual richness. This paper presents a novel retrieval augmented generation (RAG) approach tailored to descriptive football analytics, which leverages spatio-temporal and opponent-related data to transform structured event and player data into actionable insights. Our approach was evaluated using a subset from the 2023/24 season of the first German division (1. Bundesliga) over multiple game weeks, achieving an average accuracy of 63.3% in generating responses, setting a first benchmark. In particular, our approach demonstrated strong performance in answering temporal and spatial queries with an accuracy of 70%, while challenges in player-specific queries highlight opportunities for further refinement. These results underscore the potential of RAG to improve decision making for analysts, sports journalists, and potentially coaches by providing dynamic and query-specific insights, paving the way for advanced applications in descriptive sports analytics and interdisciplinary approaches in digital transformation.

The coronavirus disease 2019 (COVID-19) has had and continues to have a major impact on public health worldwide. Therefore, early detection of COVID-19 is of great importance to control the spread of the pandemic. In the course, several Machine Learning (ML) and Deep Learning (DL) approaches using various imaging modalities such as X-ray, CT, or ultrasound images have been introduced to enable faster detection and better decision making. In this context, this paper provides an overview of the state of the art (SOTA) and the development of ML and DL in the detection of COVID-19. Based on a Systematic Literature Review (SLR), a comprehensive and systematic tabular was created with the key aspects of the identified articles, such as the image modality, the techniques used for recognition, the dataset, the number of images, and the performance metrics of the developed approach. In addition, the advantages and disadvantages of the individual approaches are discussed and the need for further research is identified, which is also intended to help fight against potential future diseases.

Audio Classification using Machine Learning (ML) techniques has gained significant importance in various domains such as speech recognition, music Classification, and environmental sound analysis. Especially in combination with Transfer Learning (TL), this is a promising technique, which is why we conduct a Systematic Literature Review (SLR) on approaches in this domain, with a focus on sound Classification for monitoring tasks, which differ significantly from speech and music Classification. Furthermore, we provide an overview of TL techniques and applications, considering different methods due to the inherent characteristics of acoustic sound data. Based on our SLR, the advantages and disadvantages of the approaches are highlighted, and further research needs are identified.

Fan sound Classification is a challenging task due to the complexity of acoustic conditions and the distinctive characteristics of different microphones. This study presents an in-depth analysis of fan sound Classification using Federated Learning (FL) across three microphone setups. We evaluate the impact of microphone variations on the performance of the five FL strategies - FedAvg, FedAdagrad, FedYogi, FedAdam, and FedMedian - and explore the potential of FL for decentralized audio Classification. Our comprehensive comparison of these strategies identifies FedYogi as the most effective, demonstrating exceptional adaptability and robustness across diverse acoustic conditions. This investigation not only sheds light on the complex dynamics between microphone variations and Classification accuracy, but also offers deep insights into the application of FL for sound-based equipment monitoring. Furthermore, our findings underscore the significant impact of microphone selection on the efficacy of FL strategies, reinforcing the need for careful consideration of hardware in the deployment of FL systems.

Convolutional neural networks (CNNs) have demonstrated outstanding results in various areas of computer vision (CV). This success has led to the possibility of using CV on ever smaller computing devices, giving rise to the research area TinyML, which enables ML tasks on, e.g. resource-constrained microcontrollers. On this basis, we extend the scope of TinyML and present an image regression task where a self-generated dataset is introduced. We compare eight different approaches with different CNN architectures and normalization methods, with the best performing model achieving an MAE of 0.54 on an ESP-32. Furthermore, the ML models used are compared in terms of their performance when used on an ESP32 and on a PC. Finally, we present open questions and further research directions based on our results.

We show that a smartphone is suitable as a sensor box for processing airborne sound data in an industrial environment for predictive maintenance methods. By means of an experiment, we illustrate that our method achieves high prediction accuracies despite strong compression and can keep up with the use of an USB microphone connected to a commercial notebook. The smartphone brings a variety of advantages in the predictive maintenance context and thus surpasses conventional edge device deployments.

In a dynamic world, data streams are continuously generated, which poses immense challenges for machine learning (ML) algorithms to adapt to changing statistical properties that are subject to a non-stationary context. The underlying scenario is defined as concept drift (CD), where changes in the relationship between response and prediction variables (real CD) or a change in input data (virtual CD) are accompanied by a significant degradation in the predictive performance of the models, causing ML models to reach unacceptable levels of system accuracy. In this paper, the state of the art for CD algorithms is analyzed and compared. For this purpose, a systematic literature review was performed. Then, the 10 most popular CD algorithms were extracted from the literature using a newly-developed metric. Subsequently, the algorithms were analyzed and compared with respect to their functionality and limitations. Based on these, the optimization potentials were systematically derived. This work presents a summarized overview of CD algorithms and provides the basis for algorithm optimization in this domain.

Using the Random Forest method, we developed a fast-high-performance classification model, which can exclude a potential schizophrenic disorder in a diagnosis of potentially exposed people. Our model mainly consists of three preprocessing steps: ICA, Spectral Analysis using Buettner et al.'s 99-frequency-band-method and normalization. Using this preprocessing pipeline followed by a Random Forest, validated with different parameters, random states and a 10-fold-cross-validation, we could exclude schizophrenia with an accuracy of 100%. By applying this model in combination with a differential diagnoses system, treatments in ICUs can be done much faster, more accurately and be less expensive.