Conference Agenda

This research aims to advance medical imaging analysis by developing a novel deep learning framework for MRI segmentation. Building on the success of the BiFTransNet model, originally designed for gastrointestinal (GI) tract segmentation, we propose an enhanced architecture incorporating residual multi-scale skip connections. This design leverages the global contextual awareness of Transformer-based models alongside the localization capabilities of U-Net structures to address the inherent complexity of medical image segmentation.

The model employs a 2.5D input strategy, combining adjacent MRI slices to capture both spatial and contextual features, which significantly improves segmentation accuracy. We optimized key hyperparameter—most notably, a larger batch size—to enhance both training stability and overall performance. These enhancements resulted in marked improvements over the original BiFTransNet model.

Our proposed framework demonstrates superior performance in accurately delineating anatomical structures critical to radiation therapy planning in GI cancers.

Preoperative planning plays a critical role in Cranio-Maxillofacial surgery, where an accurate definition of osteotomy planes is essential to ensure optimal functional and aesthetic outcomes. While Virtual Surgical Planning (VSP) systems have improved process digitalization, most approaches still rely on bi-dimensional interfaces that limit spatial understanding and intuitive interaction with 3D data.

To address these limitations, this work presents an immersive Virtual Reality framework integrating multi-modal interaction to support preoperative planning tasks and enabling direct manipulation of patient-specific models and definition of osteotomy planes.

An experimental study evaluated the effect of different interaction modalities on users performance. Ten synthetic mandibular lesion cases were generated, and expert-defined osteotomy planes were used as reference that participants were asked to reproduce using both VR controller- and haptic-based interaction modalities, with performance assessed in terms of geometric accuracy and task completion time.

Results show that the haptic modality increases completion time while reducing extreme angular errors and improving consistency, with a trend toward improved angular accuracy. These findings underline the impact of interaction design on VSP, suggesting that haptics can enhance reliability in critical tasks despite increased execution time, contributing to the development of more effective immersive planning tools.

The use of surgical cutting guides for mandibular osteotomies is a consolidated approach in maxillo-facial surgery, typically based on patient-specific designs relying on anatomical fitting. However, these solutions may be limited by intraoperative variability and reduced flexibility.

The present study proposes and experimentally investigates an AR-guided hybrid cutting guide combining a physical constraint with augmented reality support for guide positioning and alignment.

A full-factorial Design of Experiments was defined to evaluate the influence of key design parameters on system performance, including the tracking modality for pose estimation (marker), the adjustment strategy of the guide (adjustment), and the type of physical constraint provided to the cutting instrument (guide).

A total of 27 configurations were tested with 40 replications each on mockups replicating irregular bone-like surfaces, considering geometric deviation, quantified through the Hausdorff distance between the planned cutting profile and the executed cut, and total task time as response variables.

The results show that the configuration combining a cubic marker, rigid adjustment, and linear guide provides the best performance, while more complex adjustment strategies and geometries lead to increased variability and execution time.

A desirability-based multi-response optimization confirmed this configuration as the optimal solution, with a global desirability index superior to all remaining alternatives.

Background. Cranial morphological assessment is crucial for early diagnosis of skull deformities in newborns, such as craniosynostosis. Typically, severity is evaluated through clinical analysis and imaging techniques, although growing interest focuses on radiation-free 3D surface scanning systems for external cutaneous evaluation. Objective. This study investigates the correlation between skin and underlying skull anthropometric measurements in pediatric patients, assessing the clinical applicability of 3D skin surface use for cranial assessment in both healthy and pathological cases. Method. 70 3D skull and soft tissue meshes were derived from CT scans of newborns (35 healthy, 35 craniosynostosis) acquired at Meyer Children’s Hospital IRCCS. Morphometric parameters including maximum length and width, diagonals at 30° and 150° and cephalic index were extracted from both surfaces and compared through a correlation analysis. Conclusion. The analysis demonstrates significant correlation between cranial and skin measurements, with Pearson correlation coefficients exceeding 0.9 across all parameters, including pathological patients. The systematic differences observed between cutaneous and cranial measurements are consistent with soft tissue thickness. These findings support the clinical applicability of surface scanning for the quantitative assessment of cranial head shape and validate its use as a reproducible, radiation-free alternative to traditional imaging techniques.

Intraosseous (IO) access is a life-saving procedure in pediatric emergencies. However, existing training tools either lack realism and durability or provide high fidelity at prohibitive costs, limiting accessibility and repeated use in routine training. A high-fidelity pediatric full-leg manikin for proximal tibial IO access was developed using CT-based anatomical modeling, modular CAD design, 3D printing, and silicone casting. A low-cost, replaceable Car-tridge with an integrated liquid vial enables realistic feedback while minimiz-ing consumable costs. Validation with 72 clinicians showed high perceived realism (mean score 6.0/7) and strong educational value, with accurate tactile feedback and fluid aspiration. The proposed simulator offers a realistic, du-rable, and cost-effective solution that supports repeated high-quality pediatric IO training.

This work presents an integrated workflow for whole-heart segmentation, animation, and visualization from 4D CT data. It combines segmentation, mesh post-processing, frame-based animation, and immersive visualization within a single web-based framework. 20 CT volumes, each corresponding to a different phase of the cardiac cycle, were segmented into the main cardiac structures using a hybrid manual and semi-automatic approach. The resulting models were refined to correct defects, discontinuities, and surface irregularities while preserving consistency across frames. Cardiac motion was reconstructed through a frame-based animation strategy, where each mesh represents a discrete temporal state of the heart. This approach avoids complex deformation methods while preserving anatomical fidelity and real-time performance.The visualization system was implemented using Babylon.js and WebXR, enabling interactive exploration of the animated model on both desktop systems and virtual reality devices. The proposed framework supports the inspection of heart morphology and its temporal evolution throughout the cardiac cycle through multiple visualization modalities, thereby enabling possible applications in diagnosis, treatment planning, and patient-specific analysis.

Traditional fall detection systems are mostly limited to binary classification, failing to describe event dynamics or severity. Large Multimodal Models (LMMs) can perform detailed semantic and biomechanical analysis from video, but their high computational demands limit continuous real-time deployment. This study presents a proof-of-concept hybrid architecture fusing a low-power wearable sensor trigger with a two-stage cloud-based LMM video analysis pipeline to characterize falls and reduce false alarms. Upon wearable detection of a candidate event, a continuously recording camera saves a synchronized video clip. A first LMM stage pre-screens the clip to filter false positives; if a fall is confirmed, a second LMM stage extracts structured fall-related parameters (initiating cause, fall direction, protective hand response, and impacted body parts) to support caregiver alerting. Validation of the pre-screening stage on 79 clips spanning three simulated fall types and three confounding activity classes yielded 87.3\% overall accuracy (precision $= 80.6\%$, recall $= 90.6\%$, specificity $= 85.1\%$). In Stage~2, evaluated on 28 real-world surveillance videos, the LMM achieved 96.4\% accuracy for initiating cause, 85.7\% for protective hand response, 71.4\% for fall direction, and a mean Jaccard index of 0.685 for impacted body-part identification. The results demonstrate the potential of integrating wearable fall detection with LMMs for structured fall characterization, although residual misclassification errors require further investigation.