Machine Learning for Biomedical Imaging

Teaching materials for the MLBI course — lecture slides, interactive dashboards, and demonstration tools for each module.
Author

Alexandre Perera Lluna, Joana Gelabert

Course materials for Machine Learning for Biomedical Imaging (MLBI), taught at UPC Barcelona and developed at B2SLab / IRIS-UPC.

Each module below links to the lecture presentation and any interactive dashboards built to accompany it. Dashboards open in the browser and require no installation.


Module 1 — Introduction to Neural Networks

Biological inspiration for neural networks, the perceptron model, activation functions, loss landscapes, and the basics of learning. Covers the historical context from Hebbian learning to modern multi-layer architectures.

NoteLecture slides

Module 2 — Multilayer Perceptrons and Learning

Forward and backward pass through multi-layer perceptrons, the chain rule, gradient descent, overfitting and regularisation, and model evaluation.

NoteLecture slides

Open presentation PDF

Interactive dashboards

Backpropagation Explorer

Step through the forward and backward pass of a small network. Inspect activations, gradients, and weight updates at each layer interactively.

Open dashboard


Module 3 — Convolutional Neural Networks

Convolution as a mathematical operation, the convolution theorem, 2-D convolutions for image processing, pooling, and the architecture of modern CNNs for biomedical image analysis.

NoteLecture slides

Open presentation PDF

Interactive dashboards

1-D Convolution Explorer

Explore how a 1-D convolution kernel slides over a signal, adjust kernel values, and observe the resulting feature map in real time.

Open dashboard

2-D Convolution Explorer

Apply user-defined kernels to 2-D images. See how edge-detection, blurring, and sharpening kernels transform the input.

Open dashboard

Convolution Theorem Explorer

Visualise the equivalence between convolution in the spatial domain and element-wise multiplication in the frequency domain.

Open dashboard

FFT Explorer

Decompose signals into their frequency components using the Fast Fourier Transform. Observe amplitude spectra and the effect of filtering.

Open dashboard


Module 4 — Recurrent Neural Networks and Sequence Modelling

Sequence modelling with RNNs, vanishing gradients, LSTM and GRU cells, bidirectional architectures, and biomedical applications including ECG classification, EEG seizure detection, and protein secondary structure prediction.

NoteLecture slides

Open presentation PDF

Interactive dashboards

RNN Unit — Forward Pass

Animated walkthrough of a single RNN cell processing a sequence step by step. Inspect hidden state updates at each time step.

Open dashboard

RNN Unit — Forward & Backward Pass

Extends the forward-pass animation with backpropagation through time (BPTT), showing how gradients flow (and vanish) across time steps.

Open dashboard

RNN — Protein Secondary Structure (forward pass)

A toy RNN trained to predict protein secondary structure. Follow the forward pass residue by residue and observe the predicted class at each position.

Open dashboard

GRU — Protein Secondary Structure Prediction

Same protein secondary structure task, now solved with a Gated Recurrent Unit. Compare predictions and gating behaviour against the plain RNN demo.

Open dashboard


Slides are built with Quarto and Reveal.js. Dashboards are self-contained HTML files requiring only a modern browser.