Abstract

Abstract
Autonomous warehouse drones require detection that is both accurate and spatially selective, detecting only the labels directly in front of the drone and ignoring those on neighbouring racks. Standard YOLOv8n has no built-in awareness of absolute pixel position, and architectural fixes (CoordConv, attention modules) are precluded by the ModalAI VOXL2 TFLite GPU runtime that constrains the deployment platform. We propose Geometric Channel Hijacking (GCH), a data-level technique that collapses the colour camera image to a single greyscale channel and replaces the two remaining input channels with horizontal and vertical positional gradients. We then run a controlled multi-factor ablation across three training regimes (INESC TEC's Autonomous Systems Laboratory (LSA), Volkswagen Autoeuropa (AE) production, and the LSA + AE combined multi-domain set), six training set sizes from n=50 to full, and six independent random seeds per cell, totalling 200 retrained models (150 across the data-size sweep and 50 at full N), on top of the original seed-0 runs, evaluated on three equalised 42-image test sets. At full data, GCH and an architecturally identical Target-Only Annotation (TOA) ablation are empirically equivalent across every metric we tested, showing that target-only annotation alone is sufficient to induce spatial selectivity in unmodified YOLOv8n. At low data, the multi-seed analysis reveals no GCH-favouring difference in mean spatial precision at p<0.05 in any of the 18 (regime, N) cells tested (paired t-test; two cells exhibit a small TOA-favouring gap in the AE regime), but TOA training in the single-domain LSA regime at N <= 100 exhibits a stochastic collapse failure mode that affects similar to 1 in 6 seeds (cross-seed std of similar to 33 percentage points on SpP vs. similar to 5.6 pp for GCH), which we also reproduce in AE at n=100 . GCH eliminates this collapse mode in the LSA regime and provides a similar to 6 & times; variance reduction at n=50 . An ablation with patience = 999 additionally shows that the collapse is recoverable in principle with roughly 2.4 & times; the standard training budget. A quantitative cross-camera test on a 1920 & times; 1080 Arducam 64 MP USB module (a completely different sensor and lens to the IMX412 used during training, no retraining performed, 115 manually annotated frames) reveals a second regime in which GCH beats TOA: cross-camera target recall is 82.5% for GCH against 64.7% for TOA at a matching spatial precision, a 17.8 percentage-point lead that confirms that the explicit positional prior provides additional robustness when the input visual statistics shift away from the training distribution. The cross-position split additionally confirms that spatial selectivity is not a memorised centre bias. We deploy the full pipeline on the VOXL2 at 22-25 FPS within the DRIVOLUTION project.

Abstract
Federated Learning (FL) enables decentralized model training while preserving data privacy but remains susceptible to poisoning attacks. Malicious clients can manipulate local data or model updates, threatening FL's reliability, especially in privacy-sensitive domains like healthcare and finance. While client-side optimization algorithms play a crucial role in training local models, their resilience to such attacks is underexplored. This study empirically evaluates the robustness of three widely used optimization algorithms: SGD, Adam, and RMSProp-against label-flipping attacks (LFAs) in image classification tasks using Convolutional Neural Networks (CNNs). Through 900 individual runs in both federated and centralized learning (CL) settings, we analyze their performance under Independent and Identically Distributed (IID) and Non-IID data distributions. Results reveal that SGD is the most resilient, achieving the highest accuracy in 87% of cases, while Adam performs best in 13%. Additionally, centralized models outperform FL on CIFAR-10, whereas FL excels on Fashion-MNIST, highlighting the impact of dataset characteristics on adversarial robustness.

Abstract
Medical image captioning is an essential tool to produce descriptive text reports of medical images. One of the central problems of medical image captioning is their poor domain description generation because large pre-trained language models are primarily trained in non-medical text domains with different semantics of medical text. To overcome this limitation, we explore improvements in contrastive learning for X-ray images complemented with soft prompt engineering for medical image captioning and conditional text decoding for caption generation. The main objective is to develop a softprompt model to improve the accuracy and clinical relevance of the automatically generated captions while guaranteeing their complete linguistic accuracy without corrupting the models' performance. Experiments on the MIMIC-CXR and ROCO datasets showed that the inclusion of tailored soft-prompts improved accuracy and efficiency, while ensuring a more cohesive medical context for captions, aiding medical diagnosis and encouraging more accurate reporting.

Abstract

Abstract
This work introduces the MassSeg-Framework, a fully automatic two-stage pipeline for breast mass analysis in mammography that integrates YOLOv11-based detection with Chan-Vese ACM refinement to achieve accurate mass localization and segmentation with a lightweight computational footprint. The framework was trained and evaluated on two publicly available datasets using consistent experimental protocols. In the detection stage, YOLOv11-nano was the most effective architecture, with a confidence threshold of 0.4, achieving statistically significant mAP50 values of 0.862 and 0.709 on the dINbreast and dCBIS datasets, respectively. These results confirm that a moderate threshold preserves clinically relevant true-positive candidates, which is particularly important for screening-oriented settings where missed lesions are costly. In the segmentation stage, the proposed framework achieved mean DICE scores of 0.721 and 0.700 on the test sets of the same datasets, demonstrating consistent overlap with expert annotations. Compared with state-of-the-art approaches that commonly assume lesion-centered ROIs or rely on heavier backbones, the proposed pipeline addresses a more realistic scenario by performing automatic detection followed by segmentation while maintaining substantially lower computational requirements. This balance between performance and efficiency makes the MassSeg-Framework a promising tool for scalable mammography analysis, particularly in resource-constrained environments or high-throughput screening workflows that require rapid processing.