ELEC5622 Signals, Software and Health
The Complete Exam Bible · S2 2026ELEC5622 Overview
- Postgraduate
- University of Sydney
- Semester 2 2026
- 6 credit points
- Close-book final exam
ELEC5622 Signals, Software and Health is a postgraduate unit in the University of Sydney's School of Electrical and Computer Engineering, worth 6 credit points. The University describes it as a health-informatics unit on how medical imaging signals are acquired, processed and analysed.
- Assessed by Two lab reports, an in-class quiz, two group projects and a final exam
- Largest single task A two-hour close-book final exam in the formal exam period
- How to prepare Rework each chapter's calculation from a blank page, then explain it in words
- Quiz scope Weeks 1 to 7, with lecture notes allowed in the room
How ELEC5622 is assessed
| Component | Weight | Format |
|---|---|---|
| Final Exam | 50% | Written exam, close-book, two hours, held in the formal exam period; AI prohibited |
| Lab report 1 | 5% | Individual experimental design report, due 30 Aug 2026 (Week 4) |
| Lab report 2 | 5% | Individual experimental design report, due 13 Sep 2026 (Week 6) |
| Quiz | 10% | In-class Canvas quiz with in-person participation, 21 Sep 2026 (Week 8), one hour |
| Project 1 | 15% | Group experimental design assignment; outline date 11 Oct 2026 (Week 9), extended on Canvas to 18 Oct 2026 |
| Project 2 | 15% | Group experimental design assignment, due 8 Nov 2026 (Week 13) |
Weights and due dates come from the University of Sydney unit outline for Semester 2 2026 and add to 100%. The outline describes the final exam as a two-hour close-book written exam in the formal exam period that assesses learning outcomes LO2 and LO3; it does not give the number or type of questions, so check any exam notice on Canvas. A Canvas announcement moved the Project 1 deadline from 11 October to 18 October 2026. Late work loses 5% of the maximum mark per calendar day.
Current ELEC5622 dates
| Date | Item | Control |
|---|---|---|
| Sun 18 Oct 2026 | Project 1 due (extended) | Moved from 11 October by a Canvas announcement; confirm on Canvas |
| Sun 8 Nov 2026, 11:59 pm | Project 2 due | Week 13 group submission |
| Formal exam period | Final Exam | Two hours, close-book; date, time and venue on your exam timetable |
Dates are as published in the ELEC5622 unit outline for Semester 2 2026 and a Canvas announcement. Confirm exact deadlines and submission settings in the live LMS.
What ELEC5622 covers
ELEC5622 moves from how medical images are formed, through MRI, diffusion MRI, CT and PET, to how they are processed and analysed: brain preprocessing, segmentation, Bayes and SVM classifiers, PCA and FLDA, classifier evaluation and neural networks.
Medical Imaging Foundations and Image Planes
Week 1. Anatomical and functional modalities, axial, coronal and sagittal planes, pixels and voxels, DICOM metadata, sensitivity and specificity.The opening chapter sets up the vocabulary the rest of the unit relies on.
- Four jobs for medical images: diagnosing, researching, planning surgery, guiding treatment
- Anatomical versus functional imaging modalities
- Axial, coronal and sagittal planes through one volume
- Pixels, voxels and voxel-averaged values
- DICOM files and why metadata matters
- Sensitivity and specificity of a screening test
MRI Signal: Spin, Precession and Relaxation
Week 2. Why hydrogen, net magnetisation, the Larmor equation, RF excitation, T1 recovery and T2 decay, tissue brightness.This chapter explains where the MRI signal comes from. Hydrogen nuclei in water behave as tiny magnets that align with the main field and precess at the Larmor frequency.
- Nuclei with spin and charge, and why hydrogen dominates
- Net magnetisation and alignment with B0
- Precession and the Larmor equation
- RF excitation and the emitted signal
- T1 recovery and the 63% rule
- T2 decay and loss of phase coherence
MRI Pulse Sequences, Spatial Encoding and k-Space
Week 2. Spin echo, TR and TE, T1 and T2 weighting, slice selection, frequency and phase encoding, k-space and scan time.This chapter turns the MRI signal into an image. Spin echo timing, set by TR and TE, decides whether contrast follows T1 or T2.
- Free induction decay and the spin echo
- Repetition time and echo time
- Rules for T1-weighted and T2-weighted sequences
- Gradient coils and the four scanner components
- Slice selection, frequency encoding and phase encoding
- The k-space matrix, the Fourier transform and scan time
Diffusion MRI, Tensor Imaging and Tractography
Week 3. Brownian motion, b-value, DWI and ADC maps, the diffusion tensor, FA and MD, deterministic and probabilistic tractography.This chapter covers imaging of water motion. Diffusion weighting makes moving water lose signal, and two images at different b-values give the apparent diffusion coefficient in every voxel.
- Brownian motion, isotropic and anisotropic diffusion
- The b-value and diffusion weighting
- DWI brightness and the ADC map
- The diffusion tensor and its ellipsoid
- Fractional anisotropy and mean diffusivity
- Colour-coded orientation maps
Brain MRI Preprocessing and Registration Pipelines
Weeks 2 and 3, Lab 1. Bias field, denoising and normalisation, skull-stripping, MNI space, affine and deformable registration, interpolation.This chapter follows a brain MRI from raw scan to regional measurement. The lecture's preprocessing problems, geometric distortion, intensity non-uniformity, noise and arbitrary intensity units, come first.
- Why preprocessing changes study outcomes
- Intensity non-uniformity and the N3 algorithm
- Denoising and intensity normalisation
- Skull-stripping with bet and its options
- Native space, MNI space and the ICBM152 template
- Rigid, affine and deformable registration
X-ray Physics and Computed Tomography
Week 4. Photon energy and wavelength, photoelectric and Compton interactions, attenuation, CT scanner generations, resolution and dose.This chapter covers X-ray physics and computed tomography. Tube voltage sets the most energetic photon, and photons are transmitted, absorbed by the photoelectric effect, or scattered. Differential attenuation forms the image.
- Photon energy, wavelength and the 12.398 relation
- Transmission, absorption and scattering
- Photoelectric effect versus Compton scatter
- Attenuation and image contrast
- Tomography versus projection radiography
- Seven generations of CT scanner
Positron Emission Tomography
PET lecture. Positron decay and annihilation, scintillation detectors, coincidence types, sinograms, attenuation correction, FBP and OSEM.This chapter explains PET, the unit's functional modality. A positron-emitting tracer such as FDG decays, each positron annihilates with an electron, and two 511 keV photons travel in opposite directions to a ring of scintillation detectors.
- Radiotracers, positron decay and half-life
- Annihilation and back-to-back 511 keV photons
- Scintillation crystals and photomultiplier tubes
- True, scattered and random coincidences
- Spatial, energy and time resolution of detectors
- Sinograms and projection geometry
Medical Image Segmentation: Edges, Regions and K-means
Image analysis lecture, Lab 2. Gradient operators, Prewitt, Sobel and Canny, region growing and splitting, K-means clustering.This chapter covers conventional segmentation, the image-processing half of medical image analysis. Edge-based methods use gradient operators, Prewitt, Sobel and Canny, then link edge pixels. Region-based methods grow from seeds or split inhomogeneous areas.
- Segmentation versus classification
- Image gradient, edge strength and direction
- Prewitt and Sobel operators
- The five steps of Canny edge detection
- Region growing and region splitting
- Unsupervised K-means clustering
Classifying Medical Images: Bayes and SVM
Image analysis lecture, Project 1. Naive Bayes, the SVM margin, support vectors, kernels and the soft margin, feature scaling.This chapter covers the decision-making half of image analysis. Naive Bayes multiplies class-conditional likelihoods by a prior.
- Applications of medical image classification
- Bayes' theorem and the naive independence assumption
- Linear SVM, the margin and support vectors
- Hard-margin optimisation and the dual problem
- Radial basis function and polynomial kernels
- Soft margins and the C parameter
Feature Extraction with PCA, Shape Models and FLDA
Week 7. The small sample-size problem, PCA by eigendecomposition, choosing k, active shape models, Fisher discriminant analysis.This chapter addresses the small sample-size problem common in medical imaging, where features outnumber patients.
- The small sample-size problem
- Feature extraction versus feature selection
- PCA as variance maximisation and eigendecomposition
- Choosing the number of components
- Reconstruction from principal components
- Active shape models and point distribution models
Evaluating Classifiers: Cross-Validation, ROC and McNemar
Week 7. Hand-out, leave-one-out and V-fold validation, ROC curves and AUC, McNemar's test for paired classifiers.This chapter is about judging classifiers honestly. Cross-validation estimates performance on unseen data, with hand-out, leave-one-out and V-fold variants.
- Discriminability on unseen data
- Hand-out validation and its pessimistic bias
- Leave-one-out and V-fold cross-validation
- Sensitivity, specificity and thresholds
- ROC curves and the area under the curve
- McNemar's test for paired classifiers
Neural Networks: MLP Training and CNNs
Week 10. Perceptron limits and XOR, losses, backpropagation, learning rate and batches, initialisation, convolutional layers.This chapter covers the neural network foundations taught in Week 10. It explains why a single layer cannot solve XOR, why hidden layers must be non-linear, and how least-squares and cross-entropy losses are minimised by gradient descent.
- From hand-crafted features to end-to-end learning
- Linear separability and the XOR problem
- Activation functions and their derivatives
- Least-squares and cross-entropy losses
- Forward and backward phases of training
- Learning rate, batches and momentum
The first half explains how images are made: image planes and voxels, MRI physics from spin and the Larmor frequency to T1 and T2 relaxation, spin echo timing, spatial encoding and k-space, diffusion MRI with the tensor, FA, MD and tractography, X-ray interactions and the generations of CT scanner, and positron emission tomography from annihilation to sinograms and OSEM reconstruction.
The second half explains how images are analysed: preprocessing and registration of brain MRI with FSL and NiftyReg, segmentation with Sobel, Canny, region growing and K-means, classification with naive Bayes and support vector machines, dimensionality reduction with PCA, active shape models and Fisher's discriminant, evaluation with cross-validation, ROC curves and McNemar's test, and the neural network foundations for the deep-learning weeks.
The unit is practical as well as theoretical.
Two lab reports, a quiz, and two group projects make up half the grade, and the first project classifies Alzheimer's disease from 90 regional grey matter volumes. The other half is a two-hour close-book final exam.
The outline lists linear algebra, probability and basic programming as assumed knowledge, and that background shows up throughout: eigenvectors in tensors and PCA, Bayes' rule in classification, and gradients in edge detection and network training. This guide works each of those calculations with printed numbers so you can redo them on paper.
Worked example · free
Apparent diffusion coefficient from two images
- 1Divide the two signals: S0 / S = 800 / 400 = 2.
- 1Take the natural logarithm: ln 2 = 0.693.
- 1Divide by the b-value difference: D = 0.693 / 800 = 0.866 × 10⁻³ mm²/s.
- 1This is more than twice 0.4 × 10⁻³, so water moves more freely here and the voxel is brighter on the ADC map.
Key terms
- Spin echo
- A spin echo is the signal regenerated when a 180° pulse refocuses spins that dephased after a 90° excitation.
- Repetition time
- Repetition time, TR, is the interval between consecutive 90° excitation pulses in a sequence.
- b-value
- The b-value is the sequence parameter that sets how strongly the MR signal is weighted by water diffusion.
- Diffusion tensor
- The diffusion tensor is a symmetric 3 × 3 matrix that captures how water diffuses in every direction within one voxel.
- Mean diffusivity
- Mean diffusivity is the average of the three diffusion tensor eigenvalues in a voxel.
- Annihilation photon
- An annihilation photon is one of the two 511 keV gamma rays emitted when a positron meets an electron.
- Coincidence detection
- Coincidence detection records two photons reaching opposite detectors almost simultaneously as one event.
- Image registration
- Image registration is the alignment of one image to another by estimating a spatial transformation.
- Gradient magnitude
- Gradient magnitude is the edge strength at a pixel, the square root of the summed squared directional derivatives.
- SVM margin
- The SVM margin is the gap between the two hyperplanes bounding the separator, equal to 2 divided by the norm of w.
- Covariance matrix
- A covariance matrix collects feature variances and covariances; its eigenvectors define the principal components.
- ROC curve
- An ROC curve plots sensitivity against one minus specificity as a classifier's decision threshold varies.
- Convolutional neural network
- A convolutional neural network is a network whose layers apply shared local filters across the image grid.
ELEC5622 FAQ
How hard is Signals, Software and Health?
It is broad rather than deep in any one area. The unit spans imaging physics, software pipelines and machine learning, and the outline assumes linear algebra, probability and basic programming. Half the grade rides on a two-hour close-book exam, so formulas such as the Larmor equation, ADC, Bayes' rule and McNemar's statistic must be remembered, not looked up.
Steady practice with small hand calculations makes the exam manageable.
How is the unit assessed?
The Semester 2 2026 outline lists six components: a final exam worth 50%, Lab report 1 and Lab report 2 at 5% each, an in-class quiz at 10%, and two group projects at 15% each. The lab reports and projects are experimental design tasks, and the projects classify brain images with conventional and deep-learning methods.
What does the outline say about the final exam?
It is a written, close-book exam of two hours, worth 50% and held in the formal exam period, with AI prohibited. It assesses LO2, the principles of imaging techniques and 3D image formation, and LO3, applying conventional and deep-learning analysis methods. Question types and numbers are not published in the outline, so watch Canvas.
What does the Week 8 quiz cover?
The quiz announcement says all content from Weeks 1 to 7 is testable. It runs in person in the lecture room during lecture time, through Canvas on your own device; you may refer to lecture notes, but internet access and AI tools are not allowed. Attempts not completed in the room are treated as invalid.
Which software does the unit use?
The brain imaging labs use FSL tools such as bet, fast, fslmaths and fslstats, the NiftyReg registration tools and the FSLeyes viewer, run in a virtual machine or a native install. Lab 2 uses Python with scikit-learn for K-means and support vector machines, and Project 1 adds shell scripts for batch processing.
Can AI help me study for this unit?
Yes, as a study aid. Sia can walk through a Larmor or ADC calculation, a backpropagation pass or a McNemar test step by step and check your reasoning. The outline allows AI in the open assessments but prohibits it in the final exam, and the quiz rules bar it in the room, so use AI to learn methods, never to produce work you submit.
What mathematics should I be comfortable with?
The outline names linear algebra and probability. In practice that means exponentials and logarithms for relaxation and diffusion, eigenvalues and eigenvectors for tensors and PCA, conditional probability for Bayes' rule, vector norms for SVM margins, and derivatives with the chain rule for edge detection and network training.
Questions students ask in each ELEC5622 chapter
- What is the difference between anatomical and functional imaging?
- Anatomical imaging, such as X-ray, CT and structural MRI, shows where structures are and how they differ in density or water content. Functional imaging, such as PET and functional MRI, shows physiological activity like glucose metabolism, usually at coarser resolution. Chapter 01 · Medical Imaging Foundations and Image Planes
- Why does MRI image hydrogen rather than other nuclei?
- Hydrogen is the only major element in tissue that is MR sensitive, and it is also the most abundant atom in the body, mostly in water. Other MR-capable nuclei such as carbon-13 or sodium-23 are far scarcer, so their signal is weak. Chapter 02 · MRI Signal: Spin, Precession and Relaxation
- Why does a spin echo need a 180° pulse?
- After the 90° pulse, spins drift out of phase and the signal fades. The 180° pulse flips their phase order so the faster spins catch the slower ones, and they come back into phase at the echo, recovering signal that seemed lost. Chapter 03 · MRI Pulse Sequences, Spatial Encoding and k-Space
- Why are three gradient directions not enough for fibre orientation?
- A fibre can run at any angle, not just along x, y or z. Measuring along many directions and fitting a symmetric 3 × 3 tensor, which has six unknowns, lets the main eigenvector point along the true fibre axis. Chapter 04 · Diffusion MRI, Tensor Imaging and Tractography
- Why register affinely before deformably?
- Affine registration removes the large differences in position, size and orientation with few parameters. Deformable registration then refines local shape. Starting the nonlinear step from a rough alignment makes it faster and less likely to settle on a wrong match. Chapter 05 · Brain MRI Preprocessing and Registration Pipelines
- Why does bone look light on an X-ray image?
- Bone has high atomic number and density, so it attenuates strongly, mainly through photoelectric absorption at diagnostic energies. Fewer photons reach the detector behind it, which produces the light areas, while soft tissue and air transmit more and appear darker. Chapter 06 · X-ray Physics and Computed Tomography
- Why does PET need attenuation correction?
- Photons crossing more tissue are attenuated more, so uncorrected images show artefacts such as lungs and skin with higher apparent uptake than muscle, and quantitation becomes wildly inaccurate. PET/CT uses the CT image to correct. Chapter 07 · Positron Emission Tomography
- Why is Sobel preferred to Prewitt?
- Sobel weights the centre row or column twice, which adds a little smoothing to the derivative. The lecture notes Prewitt is more sensitive to noise, and Sobel is the most commonly used operator for emphasising edges. Chapter 08 · Medical Image Segmentation: Edges, Regions and K-means
- What does the C parameter do in a soft-margin SVM?
- C weights the penalty on slack variables, the amounts by which points violate the margin. A large C punishes errors heavily and fits the training data closely; a small C tolerates some errors in exchange for a wider, often more general, margin. Chapter 09 · Classifying Medical Images: Bayes and SVM
- What is the small sample-size problem?
- It arises when a study has far fewer training cases than feature dimensions. Model parameters cannot then be estimated reliably, which is why dimensionality reduction by extraction or selection comes before classification. Chapter 10 · Feature Extraction with PCA, Shape Models and FLDA
- Why is hand-out validation pessimistically biased?
- Part of the data is held back and never used for training, so the classifier learns from fewer samples than it could. Its measured error is therefore usually higher than the error of a classifier trained on all the data. Chapter 11 · Evaluating Classifiers: Cross-Validation, ROC and McNemar
- Why can't a single-layer network learn XOR?
- XOR's positive and negative cases cannot be separated by one straight line: the inequalities a single linear unit would need contradict each other. A hidden layer with non-linear activations lets the network carve a boundary that no single line can draw. Chapter 12 · Neural Networks: MLP Training and CNNs
How to study for the exam
Treat ELEC5622 as two connected halves and revise them in order. For the imaging half, build a one-page sheet per modality covering the physical effect, the signal measured, the contrast mechanism, the key formula and one limitation, then practise the calculations that recur: Larmor frequency, T1 and T2 fractions, scan time, ADC, FA and MD, photon energy and wavelength, and radioactive decay.
For the analysis half, work each method by hand on a tiny example: a Sobel response on a 3 × 3 patch, K-means on ten numbers, a naive Bayes posterior, an SVM decision value and margin, a 2 × 2 PCA, an ROC curve from ten scores, a McNemar statistic and one backpropagation pass.
Because the final exam is close-book, rehearse each formula from memory before checking it, and after every calculation write one sentence on what the number means for the image or the diagnosis, the habit every published rubric rewards. Revisit the Lab 1 pipeline and the Project 1 workflow, since they tie the two halves together.
When a topic will not stick, ask Sia to explain it another way and set you a fresh example, then solve it yourself without help.
Your AI Electrical Engineering tutor for ELEC5622
Stuck on a hard ELEC5622 question? Sia is AskSia’s AI Electrical Engineering tutor — ask any ELEC5622 Signals, Software and Health question and get a clear, step-by-step explanation grounded in how the course is actually taught and assessed. Read this whole study guide free, then take your hardest questions to Sia.
