Three practical sessions are offered during this edition of the SLEIGHT Science Event – SSE#16. These sessions are progressive, and it is recommended to attend all of them.

OBJECTIVES: To gain hands-on experience with the concepts presented in previous courses, as well as with the Python libraries and code for deep learning in medical and biomedical imaging.


EQUIPMENT AND PREREQUISITES:
• Bring your own laptop. It must be functional and have a good internet connection. It is not necessary for it to be equipped with a GPU.
• You do not need to be an expert in Python. The provided code will be functional, in notebook format, and the primary objective will be to modify it to change network behavior or perform specific analyses. These notebooks will also remain available after the end of SSE#16.
• The installation guide for the software required for these sessions will be provided (Mac/Linux/Windows). It is recommended that you follow this guide and complete the installations before the practical sessions.

CONTENT OF THE PRACTICAL SESSIONS:

• Monday 6th July (16:00-17:30) - From pixels to predictions: practical deep learning for medical and biomedical images:
In this session, you will train MLP and CNN-type networks to classify MRI images into 9 classes. This first session will allow participants to finalize some installations and familiarize themselves with metrics, the use of datasets, and common Python libraries for these tasks. For advanced participants, studies on fine-tuning and knowledge distillation concepts will be offered.

• Tuesday 7th July (16:00-17:30) - Physics-Informed Neural Networks – PINNs:
Neural networks can be used to solve ordinary and partial differential equations. In this tutorial, we will explore the construction of a neural network that solves the Burgers' equation. Secondly, when data is available, we will see how to integrate it to accelerate network training or to learn equation parameters (viscosity).

• Wednesday 8th July (15:30-17:30) - Explainability.
In this practical session, we will see how to extract from the behavior of neural networks the areas of an image that contributed to its classification. The goal is to extract a justification from the network. We will then see how to modify the input image so that the network no longer assigns it the correct class. This study of counterfactual explanations in medical imaging can prove particularly useful.