Workshop Content

Three hands-on sessions | 90 minutes each | Intermediate level

Keynote Speech

This talk explores how artificial intelligence (AI) can enhance the use of ultrasound imaging in everyday clinical practice. Ultrasound is widely used because it is safe, affordable, and noninvasive, but interpreting the images can sometimes be challenging. By applying modern AI techniques such as deep learning and machine learning, we can improve the accuracy and usefulness of ultrasound in diagnosing and managing disease. In this presentation, I will share three studies. The first shows how AI models can help distinguish between benign and malignant thyroid nodules more accurately than traditional methods. The second demonstrates how AI can predict whether breast cancer has spread to nearby lymph nodes before surgery, helping doctors make better treatment decisions. The third focuses on predicting how well patients will respond to ultrasound-guided treatment for thyroid nodules. Overall, this work highlights how AI can support clinicians in making more informed, personalized decisions and improve patient care outcomes.

Keynote Speech

Understanding how transcranial magnetic stimulation (TMS) propagates through cortical networks is critical for both basic neuroscience and clinical applications. We present neurofe, an open-source Python package that enables researchers to fit neural mass models to empirical TMS-EEG data using gradient-based optimization. Built on PyTorch, neurofe implements the Jansen-Rit neural mass model with biologically informed features including conduction delays, structural connectivity, and leadfield projection, following the framework established by Momi et al. (2022, Nature Communications). A key feature is virtual dissection — users can programmatically lesion specific brain regions and observe the resulting changes in simulated EEG, providing insights into the causal role of individual regions within whole-brain dynamics. The package offers a clean, high-level API where fitting a model requires just a few lines of code, while also exposing a standalone NumPy simulator for lightweight parameter exploration. neurofe is pip-installable, includes comprehensive documentation and ready-to-run examples, and aims to lower the barrier to entry for whole-brain computational modeling in the TMS-EEG research community.

Session 1: Hotaru

This session introduces Hotaru, a novel multi-protocol networking framework designed for simultaneous operation over HTTP, MQTT, and other protocols within a single application. Participants will explore how Hotaru simplifies the development of connected healthcare systems, where medical devices, data aggregation pipelines, and clinical dashboards must communicate across heterogeneous networks. Through three hands-on exercises; a Hello World implementation, multi-protocol exploration, and a simple device data aggregation system; attendees will gain practical experience building the connectivity layer that underpins modern connected care environments.

Session 2: G-nome

This session examines the challenge of deploying AI in clinical genomics safely and reliably. Participants will explore G-nome, a clinical genomics platform that places verified evidence at the centre of every analysis. Rather than relying on generative models alone, G-nome combines a deterministic knowledge engine with a hallucination-guarded AI reasoning layer to produce fully traceable, citation-backed findings for both clinicians and patients. Topics covered include the ACMG 5-tier variant classification framework, pharmacogenomics (PGx) guidance using CPIC evidence levels, dual-interface architecture for clinical and patient use, and regulatory considerations under Hong Kong's PDPO and the Asia-Pacific context.

Session 3: PulmoSight

This session addresses a core challenge in deploying medical AI: state-of-the-art models are too large and slow for mobile devices, edge hardware, and resource-constrained clinical settings. Participants will implement a knowledge distillation pipeline in which a large EfficientNet-B3 teacher model transfers its learned representations to a compact MobileNetV2 student, achieving competitive diagnostic accuracy at a fraction of the computational cost. Working with the SIIM-ISIC melanoma dataset, attendees will implement the distillation loss function (hard labels + soft labels with temperature scaling), run a training loop, compare teacher and student performance, and export the final model to TorchScript and ONNX formats for real-world deployment on mobile apps, web APIs, and edge devices.