Abstracts Track 2025

Area 1 - Fundamentals

Nr:	33
Title:	Predicting Particle Fate Following Phagocytosis with Explainable Machine Learning-Based Microscopy
Authors:	Rebecca Betts and Ashley Cadby
Abstract:	Phagocytosis, the engulfment and elimination of particles by cells, is a fundamental component of the innate immune response. Bacteria have evolved various strategies to evade phagocytosis, some of which are well understood, while others remain unidentified. Identifying patterns in phagocyte morphology and dynamics that indicate successful or failed phagocytosis events can help uncover such mechanisms. Using time-lapse, two-channel microscopy, we image murine J774 macrophages and Staphylococcus aureus bacteria to capture phagocytosis events. Segmentation via CellPose and tracking with a linear assignment problem approach enable extraction of multivariate time series features describing the morphology, movement, and interactions of cells involved in each event. Each series of features is classified as successful or unsuccessful phagocytosis. To classify these events and determine patterns which are indicative of phagocytosis success, we plan to evaluate several machine learning approaches, balancing accuracy and interpretability. The random forest classifier will use summary statistics derived from the feature time series, though it may not capture temporal patterns effectively. Long Short-Term Memory networks will be explored for learning complex sequential data, with attention mechanisms incorporated to indicate the importance of each time step in the final classification. Generalized additive models will be considered for their inherent interpretability, though they may limit accuracy due to their handling of feature interactions. Finally, 1D convolutional neural networks will be tested, and class activation mapping used to provide interpretability in their predictions. By applying these techniques, this work aims to develop a framework for linking phagocyte behaviour to phagocytosis outcomes and identifying bacterial evasion strategies.

Area 2 - Methods and Techniques

Nr:	41
Title:	Automatic Detection of Wave Overtopping in Coastal Videos Using Deep Learning
Authors:	Fábio André Carvalho Santos, Telmo Reis Cunha and Paulo Renato Enes Baganha Baptista
Abstract:	Coastal video monitoring systems are widely used for managing coastal zones due to their cost-effectiveness. In June 2021, a station was installed at Praia dos Pescadores, Ericeira, Portugal, capturing 10-minute videos during daylight hours, generating approximately 72 videos daily. The station monitors the northern harbor breakwater, which is prone to wave overtopping, a phenomenon that poses risks to infrastructure and causes significant local economic and environmental impacts. This study explores a Deep Learning approach, offering a more effective alternative to traditional image processing methods for timely and accurate event detection. The proposed model employs a hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) architecture to automatically detect wave overtopping events in video sequences. CNNs are highly effective at extracting spatial features from images, while LSTMs capture temporal dependencies, making this combination particularly well-suited for analyzing video sequences. This approach allows the model to process spatial and temporal information simultaneously, enhancing its ability to detect foam movement indicative of overtopping events. The dataset for this study consisted of 3058 labeled image sequences derived from 10-minute recordings, with 1584 sequences containing wave overtopping events and 1474 without. The sequences were selected to cover various types of overtopping, locations along the breakwater, times of day, and weather conditions. Each sequence was preprocessed by masking irrelevant areas, cropping to focus on the breakwater structure, and resizing the frames to 224x224 pixels to match the input size required by the CNN (GoogLeNet). After preprocessing, the model was trained on 85% of the sequences, with the remaining 15% reserved for validation. Hyperparameter tuning was performed using Bayesian optimization. The final architecture combined GoogLeNet for spatial feature extraction with a bidirectional LSTM (BiLSTM) layer containing 1480 units (740 in each direction) to capture the temporal dependencies of overtopping events. This hybrid approach allows the model to learn both the spatial dynamics of foam movement and the temporal structure of overtopping events. Performance was evaluated using a test set of 814 videos (461 with overtopping events and 386 without). The model achieved high precision (0.97), recall (0.92), specificity (0.97), and F1-score (0.94), as well as an AUC-ROC of 0.94, indicating its effectiveness in detecting overtopping events with minimal false positives. Despite the promising results, challenges were encountered, particularly in detecting overtopping during late hours, likely due to underrepresentation in the dataset. Future work should expand the dataset to include more diverse overtopping scenarios, improving model adaptability. The results suggest that this methodology could be integrated into coastal monitoring systems, enhancing hazard response and reducing risks. The CNN-BiLSTM model shows strong potential for automating overtopping detection, thereby supporting more efficient coastal management.