Dan Popescu

The detection of ovarian cancer using computed tomography images is making rapid progress, combining the strengths of different neural networks within a complex ensemble, a more effective model can be created to support medical decision-making in ovarian cancer cases. Various neural networks and ensemble models were tested on the TCGA OV database. Tumor regions were segmented using computer-assisted annotation masks, while molecular subtype labels were assigned at the patient level based on TCGA genomic classifications and mapped to tumors. As the main novelty, the proposed method involves a two-stage fusion system. The first stage combines several neural networks into four ensembles to capture diverse features, and the second stage integrates these results into a global system to enhance segmentation accuracy. The selected models demonstrated strong segmentation performance within the investigated experimental setting and were subsequently integrated into the proposed two-stage fusion framework. The neural networks for the first stage are Attention U-Net, DeepLabV3+, and ViT. The results show that the proposed two-stage fusion system achieved higher observed segmentation performance than the evaluated individual neural networks and first-stage ensemble configurations within the present TCGA-OV experimental setting. Across the five evaluated classes, the Global System achieved an average F1 score of 93.53%, recall of 93.38%, and precision of 93.68%, supporting the technical feasibility of the proposed two-stage fusion approach for ovarian tumor segmentation within the investigated dataset.

The optimization of herbicide application is one of the most important topics in Precision Agriculture, driven by both economic efficiency and ecological sustainability. Excessive herbicide use can lead to soil degradation, water contamination, and negative impacts on biodiversity, while also contributing to human health risks and climate-related concerns. Developing accurate, automated approaches for distinguishing crops from weeds is therefore essential to support sustainable agricultural practices. In this paper, a novel architecture for crops and weed segmentation in tobacco plantations is proposed: a U-Net variant which incorporates several specific design elements, including deep supervision, a Vegetation Global Context block, and a dual-headed output that separately predicts vegetation and crop masks. Weed regions are derived as the difference between vegetation and crop predictions, allowing the model to enforce logical consistency directly within a single framework, in contrast to other two-step approaches. The proposed architecture was evaluated using multiple modern encoder backbones (ConvNextV2, FastViT, RepViT, MambaVision). Experimental results demonstrate that this architecture not only improves segmentation accuracy compared to prior approaches, with best scores of 94.24% Dice for crop segmentation and 93.72% for weeds, but also significantly reduces inference time by avoiding multi-stage pipelines, making it well-suited for real-time deployment.

The optimization of herbicide application is one of the most important topics in Precision Agriculture, driven by both economic efficiency and ecological sustainability. Excessive herbicide use can lead to soil degradation, water contamination, and negative impacts on biodiversity, while also contributing to human health risks and climate-related concerns. Developing accurate, automated approaches for distinguishing crops from weeds is therefore essential to support sustainable agricultural practices. In this paper, a novel architecture for crops and weed segmentation in tobacco plantations is proposed: a U-Net variant which incorporates several specific design elements, including deep supervision, a Vegetation Global Context block, and a dual-headed output that separately predicts vegetation and crop masks. Weed regions are derived as the difference between vegetation and crop predictions, allowing the model to enforce logical consistency directly within a single framework, in contrast to other two-step approaches. The proposed architecture was evaluated using multiple modern encoder backbones. Experimental results demonstrate that this architecture not only improves segmentation accuracy compared to prior approaches, with best scores of 94.24% Dice for crop segmentation and 93.72% for weeds, but also significantly reduces inference time by avoiding multi-stage pipelines, making it much better suited for real-time deployment in field conditions.

Breast cancer is a serious oncological condition whose treatment has often been associated with the loss of the breast or both breasts. If discovered early, it can be treated and has a favorable long-term prognosis, yet manual interpretation is time-consuming. In this context, the article proposes a deep neural network pipeline for the automatic detection and precise localization of breast tumors from MRI images. The images were taken from the Duke Breast Cancer MRI Dataset and were preprocessed and organized for classification and detection using convolutional neural network architectures such as VGG16 and YOLOv9m. The VGG16 model achieves 93.77 % accuracy, and the YOLOv9m model achieves 94.7% mAP50. In addition, the VGG16 classifier achieved a precision of 95.18%, a recall of 92.34 %, and an $\mathbf{F 1}$-score of 93.73 %, while the YOLOv9m detector achieved a precision of 96.8 % and a recall of 90.1 %. These metrics enhance the clinical relevance of the obtained results and further highlight the strong performance achieved compared to other studies in the literature in the field.

Multimodal systems have shown better performance in emotion recognition than single-modality approaches because they combine different types of information that complement each other. Among the most common combinations are video and audio, which together can capture both visual and acoustic signs of emotion. In this paper, we present a multimodal architecture that uses transformer models for video processing and Mel-Frequency Cepstral Coefficients (MFCC) frames as time-series input for the audio stream. The proposed system classifies video samples from the CREMA-D dataset into six emotion categories: Anger, Disgust, Fear, Happy/Joy, Neutral, and Sad. To understand the contribution of each modality, video and audio were first analyzed separately. The video transformer was effective in capturing motion and appearance patterns, while the MFCC-based model extracted useful audio features that yielded good classification results on their own. After combining them, the multimodal system demonstrated higher accuracy than either modality alone. These results confirm that integrating video and audio provides a stronger and more reliable solution for emotion classification.

Unmanned aerial vehicles (UAVs) are increasingly deployed in critical environments (e.g., borders, airports), where real-time detection and tracking remain challenging due to small target size, complex backgrounds, long ranges, and resource constraints. We present DeTrAck, a non-invasive UAV monitoring system designed for embedded platforms that couples a YOLO11-based detector with an in-loop validation convolutional neural network (VaL_CNN) and DeepSORT tracking, while issuing pan-tilt-zoom (PTZ) commands via ONVIF to maintain target on-sight. We curated a custom dataset comprising both visible spectrum and infrared imagery. Across recorded sequences, the proposed ensemble reduces false positives, enables lower detection thresholds, and sustains continuous tracking. In a 3,647-frame sequence, it preserved target identity without loss, achieving competitive detection performance (≈ mAP@50 78%) at near real-time throughput (≈10% FPS reduction versus baseline). An ablation study across YOLO11 variants shows that the YOLO11s + VaL_CNN configuration provides the best precision–speed trade-off, resulting in an over 4% increase in true positives and a 2.6 times inference processing reduction in false positives compared to the baseline detector even for very small targets 1:2500. DeTrAck supports multi-target and multi-camera operation via RTSP/ONVIF integration, making it suitable for practical surveillance deployments on portable ground-air observation points.