Deep Learning Fundamentals and Health Applications

 02/04/2025 (11-13)

Deep Learning Fundamentals I

deep neural networks; gradient issues; activation functions; normalization and regularization; optimization

Deep Learning Fundamentals II

neural autoencoders; unsupervised deep learning; autoencoding tasks in healthcare (anomaly detection, compression, denoising)

[6] Survey on autoencoders

[7,8] Surveys on autoencoders use in healthcare

Convolutional Neural Networks I

Introduction to medical imaging; basic CNN elements; 

Convolutional Neural Networks II

CNN architectures; medical imaging tasks

[SD] Chapter 10

[AI4H] Pg. 643-655

Additional readings

[9] Augmentation techniques for medical imaging

[10] Survival analysis with CNNs

[11] Large-scale medical image segmentation dataset and application

Software

nnU-net - Well engineered framework for low-code medical segmentation tasks

Lab 8: Prova in itinere

AI Meets Psychiatry: fMRI-Based Multi-Disorder Diagnosis - Guest Lecture by Elisa Ferrari (CEO, Quantabrain)

QuantaBrain is a startup developing AI models for the diagnosis and characterization of psychiatric disorders using a short resting-state functional MRI scan, a complex 4D dataset influenced by numerous variables. In this seminar, the QuantaBrain team will present their neural network architecture, which combines adversarial learning with optical flow. They will share the technical challenges they encountered while training this network and scaling it from one to eleven disorders. Finally, participants will get an exclusive preview of the research platform they are currently building, designed to support scientists working on psychiatric disorders.

Lab Tutorial 9: Deep Neural Networks for image classification

Lab Tutorial 10: Deep Neural Networks for image segmentation

Spring Break: No Lectures

Deep learning for sequential data I

definitions; sequential data in healthcare; physiological timeseries; electronic health records; recurrent neural networks; learning long-term relationships and gradient issues

Coverage of  course books on this lecture is inadequate. You can use the course slides for this topic, and if you like you can integrate those with chapter 10 from the Deep Learning Book.

You can skip Sections 10.2.2, 10.2.3, 10.2.6)

Physionet- the reference repository for biomedical signals/timeseries

MIMIC-IV - the reference EHR dataset (in its latest version), also part of Physionet

Deep learning for sequential data II

gated recurrent networks; LSTMs and GRUs; bidirectional models; convolutional RNNs; applications of RNNs in healthcare

Coverage of  course books on this lecture is inadequate. You can use the course slides for this topic, and if you like you can integrate those with chapter 10 from the Deep Learning Book.

You can skip Sections 10.2.2, 10.2.3, 10.2.6)

Lab Tutorial 11: RNNs

Encoder-Decoder  Architectures and Neural Attention

sequence-to-sequence learning,  encoder-decoder architectures, cross-attention; self-attention; transformer architectures

Deep learning with textual sequences I

representing textual information, word embeddings, skip-grams, tackling textual modelling tasks, masked language modelling

[SD] Chapter 12

[AI4H] Pg. 623-636

Deep learning with textual sequences II

relevant language model architectures, pretraining and fine tuning, language models in healthcare, foundation models

Lab Tutorial 12: LLMs