Balazs Harangi

Domain shift between microscopy imaging domains poses a significant challenge for deploying deep learning-based cell detection models across different experimental setups. Manual annotation of new microscopy datasets remains resource-intensive and time-consuming. This study presents a detection-oriented Cycle-Consistent Generative Adversarial Network (CycleGAN)-based annotation transfer framework for adapting a labeled B16BL6 source domain to a HeLa target domain. Annotated B16BL6 melanoma microscopy images are translated into the visual appearance of HeLa microscopy data while retaining their original bounding-box annotations, enabling YOLOv8x detector training without full manual annotation of the target domain. Three YOLOv8x configurations were compared: a source-only bright B16BL6 baseline, an intensity-inverted dark B16BL6 baseline, and the proposed CycleGAN-translated B16BL6 → HeLa configuration. Performance was evaluated using standard object detection metrics on a manually annotated 100-frame HeLa target-domain subset, together with complementary unsupervised proxy metrics on the full unlabeled HeLa dataset. The CycleGAN-trained detector achieved the highest supervised target-domain performance, with a precision of 0.244, recall of 0.353, F1-score of 0.288, and mAP@0.50 of 0.186, compared with mAP@0.50 values of 0.027 and 0.009 for the bright and dark baselines, respectively. It also achieved the highest exploratory composite reliability score on the full HeLa sequence. These findings demonstrate that source-to-target image translation improves detector generalization under the investigated B16BL6 → HeLa phase-contrast microscopy domain shift, thereby reducing the need for extensive manual target-domain annotation.

In this paper, a novel approach is presented for fusing RGB and LiDAR inputs for semantic segmentation. Accurate building detection is required for various scenarios such as urban planning or environmental monitoring. The two main sources for accurate building segmentation are either RGB aerial images or LiDAR point clouds covering the selected area. Each of these sources has its own well-known techniques for segmentation; however, for the combination of the input, there are not many architectures available, and extracting different features from the two different fields can result in an enhanced segmentation map. The authors of this article created a semantic segmentation model that uses both the aerial RGB image and the LiDAR point cloud as its input. The network first takes the point cloud and forwards the processed projection to a modified U-Net-based architecture, which fuses the extracted features of the 3D input with the extracted information of the 2D input on each level of the decoding. To train and test the presented model, the authors used a dataset containing more than 3000 images and their corresponding 3D point clouds of three different areas from Hungary. As is also presented in this paper, this approach provides significantly better results than the traditional RGB, Point Cloud segmentation models, and their ensembles in terms of segmentation accuracy.

Accurate and automated detection and tracking of cells in microscopy images is a persistent challenge in biotechnology and biomedical research. Effective detection and tracking are crucial for understanding biological processes and extracting meaningful data for subsequent simulations. In this study, we present an integrated pipeline that leverages a fine-tuned YOLOv8x model for detecting cells and cell divisions across microscopy image series. While YOLOv8x exhibits strong detection capabilities, it occasionally misses certain cells, leading to gaps in data. To mitigate this, we incorporate the DeepSORT tracking algorithm, which enhances data association and reduces the cells’ identity (ID) switches by utilizing a pre-trained convolutional network for robust multi-object tracking. This combination ensures continuous detection and compensates for missed detections, thereby improving overall recall. Our approach achieves a recall of 93.21% with the enhanced DeepSORT algorithm, compared to the 53.47% recall obtained by the original YOLOv8x model. The proposed pipeline effectively extracts detailed information from structured image datasets, providing a reliable approximation of cellular processes in culture environments.

The Morris Water Maze (MWM) is a widely used behavioral test to assess the spatial learning and memory of animals, particularly valuable in studying neurodegenerative disorders such as Alzheimer’s disease. Traditional methods for analyzing MWM experiments often face limitations in capturing the complexity of animal behaviors. In this study, we present a novel AI-based automated framework to process and evaluate MWM test videos, aiming to enhance behavioral analysis through machine learning. Our pipeline involves video preprocessing, animal detection using convolutional neural networks (CNNs), trajectory tracking, and postprocessing to derive detailed behavioral features. We propose concentric circle segmentation of the pool next to the quadrant-based division, and we extract 32 behavioral metrics for each zone, which are employed in classification tasks to differentiate between younger and older animals. Several machine learning classifiers, including random forest and neural networks, are evaluated, with feature selection techniques applied to improve the classification accuracy. Our results demonstrate a significant improvement in classification performance, particularly through the integration of feature sets derived from concentric zone analyses. This automated approach offers a robust solution for MWM data processing, providing enhanced precision and reliability, which is critical for the study of neurodegenerative disorders.

Machine learning-based systems have become instrumental in augmenting global efforts to combat cervical cancer. A burgeoning area of research focuses on leveraging artificial intelligence to enhance the cervical screening process, primarily through the exhaustive examination of Pap smears, traditionally reliant on the meticulous and labor-intensive analysis conducted by specialized experts. Despite the existence of some comprehensive and readily accessible datasets, the field is presently constrained by the limited volume of publicly available images and smears. As a remedy, our work unveils APACC (Annotated PAp cell images and smear slices for Cell Classification), a comprehensive dataset designed to bridge this gap. The APACC dataset features a remarkable array of images crucial for advancing research in this field. It comprises 103,675 annotated cell images, carefully extracted from 107 whole smears, which are further divided into 21,371 sub-regions for a more refined analysis. This dataset includes a vast number of cell images from conventional Pap smears and their specific locations on each smear, offering a valuable resource for in-depth investigation and study.

A simple and cheap way to recognize cervical cancer is using light microscopic analysis of Pap smear images. Training artificial intelligence-based systems becomes possible in this domain, e.g., to follow the European recommendation to screen negative smears to reduce false negative cases. The first step for such a process is segmenting the cells. A large and manually segmented dataset is required for this task, which can be used to train deep learning-based solutions. We describe a corresponding dataset with accurate manual segmentations for the enclosed cells. Altogether, the APACS23 (Annotated PAp smear images for Cell Segmentation 2023) dataset contains about 37 000 manually segmented cells and is separated into dedicated training and test parts, which could be used for an official benchmark of scientific investigations or a grand challenge.

Chest X-rays are vital in healthcare for diagnosing various conditions due to their low Radiation exposure, widespread availability, and rapid interpretation. However, their interpretation requires specialized expertise, which can limit scalability and delay diagnoses. This study addresses the multi-label classification challenge of chest X-ray images using the Chest X-ray14 dataset. We propose a novel online ensemble technique that differs from previous penalty-based methods by focusing on combining individual model losses with the overall ensemble loss. This approach enhances interaction and feedback among models during training. Our method integrates multiple pre-trained CNNs using strategies like combining CNNs through an additional fully connected layer and employing a label-weighted average for outputs. This multi-layered approach leverages the strengths of each model component, improving classification accuracy and generalization. By focusing solely on image data, our ensemble model addresses the challenges posed by null vectors and diverse pathologies, advancing computer-aided radiology.