1
Q
What is the Fourier Series?
A
The Fourier Series is a way to represent a periodic function as a sum of sine and cosine waves.
- It breaks down complex periodic signals into simpler trigonometric components.
- Each term in the series has a specific frequency, amplitude, and phase.
- Used in signal processing, physics, and engineering to analyse waveforms.
2
Q
What are the key steps in analysis?
A
- Identify problem
- Find dataset
- Get ethics approval
- Check data acquisition & consent
- Ensure anonymisation
- Decide metrics & ground truth
- Start analysis & problem solving
3
Q
What is the difference between analogue and digital video?
A
Analogue video → uses continuous signals (like VHS tapes). Quality drops with noise or copying.
Digital video → uses binary data (0s and 1s). Quality stays consistent, easy to edit, supports HD/4K/8K.
4
Q
What is a sinusoid?
A
A sinusoid is a smooth, repetitive wave that describes the shape of sine or cosine functions.
- It oscillates between a maximum and minimum value.
- Commonly used to model periodic phenomena like sound waves, light waves, and alternating current.
- Defined by its amplitude, frequency, phase, and offset
5
Q
What is frequency?
A
f = 1/T
- how many cycles per second.
- T is the time taken to complete one wave cycle
6
Q
What is phase (theta)?
A
Essentially how much the wave moves left or right. If theta = 0, the wave has not moved.
Generally an exam question will not be in context of a phase shift.
7
Q
What is the equation of a sinusoid?
A
𝑦 (𝑡) = 𝐴 ∙ sin(𝜔 ∙ 𝑡 + 𝜃)
A = amplitude
T = time for one wave completion
f = frequency = 1/T
𝜔= angular frequency= 2𝜋f= 2𝜋/T
𝜃 = phase difference
8
Q
What is the fourier series?
A
Any periodic signal s(t) may be represented as the sum of a set of sinusoidal waves of different frequencies and phases.
- any signal can be described as the
sum of sinusoidal waves of different frequencies.
9
Q
What is the highest an analog transmission system will transmit?
A
450Hz.
10
Q
What is the maximum of a digital transmission system bit rate?
A
4kbps, 8bit samples
11
Q
What is the frequency domain?
A
The frequency domain is a way of analysing signals based on the frequencies they contain, rather than how they change over time.
- Transforms time-based signals into frequency components using tools like the Fourier Transform.
- Reveals how much of each frequency is present in a signal.
12
Q
What is the Fast Fourier Transform (FFT)?
A
The FFT is an efficient algorithm to compute the Discrete Fourier Transform (DFT) of a signal.
- Converts a signal from the time domain to the frequency domain.
- Reveals the frequency components present in a discrete signal.
- Reduces computational complexity from O(n^2) to O(n \log n), making it ideal for real-time processing.
13
Q
What are some applications of FFT?
A
- Audio and image compression
- Signal filtering
- Spectral analysis
- Radar and communications
14
Q
Where does the square shape of a square wave come from?
A
The summation of many frequencies.
15
Q
What is the spectrum of a signal?
A
- A mathematical transformation that represents all the information of the signal.
- From the spectrum we can get back all the signal in the time domain (aka lossless)
- if we compressed the signal, it would now lose info (lossy)
16
Q
What is the difference between the Fourier Series and Fourier Transform?
A
The Fourier series is used to represent a periodic function by a discrete sum of complex exponentials, while the Fourier
transform is then used to represent a general, nonperiodic function by a continuous superposition or integral of complex exponentials.
17
Q
What is a low-pass filter?
A
A low-pass filter allows low-frequency signals to pass through while attenuating high-frequency signals.
- Smooths out rapid changes or high-frequency noise.
- Used in audio to reduce hiss or sharp sounds.
- Essential in data smoothing and anti-aliasing in graphics.
18
Q
What is frequency filtering?
A
Frequency filtering is the process of modifying or removing specific frequency components from a signal.
Types of Filters:
- Low-pass: Keeps low frequencies, removes high.
- High-pass: Keeps high frequencies, removes low.
- Band-pass: Keeps a specific range of frequencies.
- Band-stop (notch): Removes a specific range of frequencies.
19
Q
What is a high-pass filter?
A
A high-pass filter allows high-frequency signals to pass through while attenuating (reducing) low-frequency signals.
- Removes slow or low-frequency components (e.g., background hum).
- Used in audio processing to eliminate bass or rumble.
- Common in electronics, signal processing, and image enhancement.
20
Q
A signal is only the variation of one variable (e.g., voltage) in time.
A. Yes
B. No
C. I don’t know
A
B. No
21
Q
If you know the frequency of a signal, you also know its angular frequency and its period.
A. Yes
B. No
C. I don’t know
A
A. Yes
22
Q
A sinusoidal signal is completely characterized by its amplitude and frequency.
A. Yes
B. No
C. I don’t know
A
B. No
23
Q
A Fourier Transform is an alternative way to represent a signal.
A. Yes
B. No
C. I don’t know
A
A. Yes
24
Q
Electromagnetic radiation with wavelengths between 400nm and 700nm is known as?
A
Visible light /visible spectrum (400nm is the top visible frequency, frequency decreases with wavelength)
25
Each cone cell can distinguish a large number of different frequencies.
A. Correct
B. Incorrect
C. Depends
B. Incorrect
26
What are vector images?
Higher level representations of images as a combination of mathematical descriptions of curves, shapes, gradients.
- still need to be transformed into raster for displays and printers which can lead to a different interpretation by the rasteriser
- type of source encoding
- use for tables and graphs
27
What are the pros and cons of using a bitmap?
Pros:
* Stable
* Fits most presentation media (e.g.,
displays)
* Fits many capture media (e.g., CCD
capture)
* Can use lossy compression
Cons:
* Not Scalable
* Difficult to animate
* Bad for 3D (voxels are expensive)
* Limited spatial frequency
– Results in aliasing
28
What are the pros and cons of using a vector image?
Cons:
* Render-dependent
* Large variability in encoding (# of
objects)
* Might encode invisible information
* Non-realistic
* Lossy compression generally not
possible
Pros:
* Scalable
* Potentially efficient
* Easy to animate
* Good for 3D
* More possible transformations
29
What is SVG?
- Scalable Vector Graphics
- Supported (in varying degrees) natively by most current browsers
- source compression as it is XML
30
What are some basic elements of SVGs?
* Paths (curved or straight lines)
* Basic Shapes
* Text
* Painting (filling, outlines, strokes, gradient, pattern)
* Colour (rgb)
* Gradients and Patterns
* Clipping, Masking and Compositing
* Filter Effects
* Interactivity (events, focus)
* Linking
* Scripting
* Animation
* Fonts
* Metadata
31
How might you reduce quantisation noise?
- reduce with higher bit depth
32
How might you reduce thermal noise?
- reduce by cooling
33
How can you reduce background signals?
- reduce by filtering
34
How can you reduce noise from lossy encoding?
- reduce by using lossless or less lossy encoding
35
What is the difference between Discrete and Digital Signals
Discrete Signal
- Defined only at specific time intervals
- Can have any value (continuous amplitude) at those intervals
- Example: Temperature readings every hour
Digital Signal
- Defined at specific time intervals and
- Has finite, quantized values (usually binary: 0s and 1s)
- Example: Computer data transmission
36
What is aliasing?
Aliasing happens when converting a continuous signal (like sound or light) into a discrete digital signal through sampling, and the sampling rate is too low to accurately represent the original signal.
37
What is the Nyquist-Shannon Sampling Theorem?
Sample at at least double the maximum frequency of your signal.
38
The bandwidth of a signal is…
A. The list of the frequencies that it can be
decomposed into
B. The value of its highest frequency
C. How much width it occupies in the frequency
domain
D. The difference between its highest frequency
and its lowest frequency
E. C and A
F. B and D
G. C and D
g) C and D
39
The bandwidth of the channel always has to be larger than the bandwidth of the signal.
A. Yes
B. No
C. It depends
D. I don’t know
B. No
C. It depends
40
What are some dB examples? E.g. 0dB is barely audible...
* 0dB – barely audible sound (more later)
* 20dB – Whisper
* 40dB – Quiet office
* 60dB – Normal conversation
* 80dB – Hair dryer
* 100dB – Heavy traffic, pneumatic drill
* 120dB – Loud thunder, music concert
* 140dB – Jet aircraft at take off
41
How does human hearing work (step 1: outer ear)?
1. Outer Ear: Capturing Sound
- The pinna (the visible part of the ear) collects sound waves from the environment.
- These waves travel through the ear canal and strike the eardrum (tympanic membrane), causing it to vibrate.
42
How does human hearing work (step 2: middle ear)?
2. Middle Ear: Amplifying Vibrations
- The vibrations from the eardrum are transferred to three tiny bones called the ossicles:
- Malleus (hammer)
- Incus (anvil)
- Stapes (stirrup)
- These bones amplify the sound and transmit it to the oval window, a membrane leading to the inner ear.
43
How does human hearing work (step 3: inner ear)?
3. Inner Ear: Converting to Electrical Signals
- Vibrations enter the cochlea, a fluid-filled, spiral-shaped structure.
- Inside the cochlea, vibrations create waves in the fluid, which move the basilar membrane.
- Hair cells on this membrane bend in response, opening channels that generate electrical signals.
44
How does human hearing work (step 4: Auditory Nerve & Brain)?
4. Auditory Nerve & Brain: Interpreting Sound
- The auditory nerve carries these electrical signals to the brainstem and then to the auditory cortex in the temporal lobe.
- The brain processes these signals, allowing us to recognize and understand sounds like speech, music, or environmental noise.
45
What are some physical measures of sound (sound perception)?
Physical Measures:
* Intensity (amplitude)
* Frequency
* Spectrum (complexity)
46
What are some perceptual measures of sound (sound perception)?
Perceptual Sensations:
* Loudness
* Pitch
* Timbre
47
What is Masking (perceptual irregularity)?
The rise in the detection threshold of one tone (test tone) due to the presence of a second tone (masker tone)
48
What is temporal masking?
Temporal Masking (Time-Based)
- Definition: A sound becomes inaudible because another sound occurs immediately before or after it.
- Types:
- Forward masking: The masker comes before the target sound.
- Backward masking: The masker comes after the target sound.
- Example: A loud drumbeat can make a soft whisper just before or after it harder to hear.
- Mechanism: Related to how the auditory system processes sounds over time—there’s a brief window where sounds can interfere with each other even if they don’t overlap.
49
What is Spectral Masking?
Spectral Masking (Frequency-Based)
- Definition: A sound becomes inaudible because another sound at a similar frequency occurs simultaneously.
- Example: A loud tone at 1000 Hz can mask a quieter tone at 1050 Hz if played together.
- Mechanism: Occurs due to the way the cochlea processes overlapping frequencies—sounds close in frequency activate similar regions, making it harder to distinguish them.
50
Which of two pure sinusoids, 300Hz and 3kHz, of 20 phons sounds louder?
Which one will have a larger amplitude?
Phons are already corrected for the irregular human perception in different frequencies, so they will sound approximately equally loud.
To compensate for
the less sensitive perception at lower frequencies, the
300Hz signal has to be louder (aka larger amplitude).
51
What is dynamic range?
Ratio between the largest and smallest differentiable signals.
52
How do microphones work?
1. diaphram moves due to sound wave
2. vibration becomes electrical signal
3. signal travels to gear
4. gear boosts signal
5. signal turned back into sound so we can hear it (speaker)
-> type of transducer -> converts waves to signals
53
What happens if the music we are sampling has components above 22kHz?
We filter them out!
Otherwise they can create undesired frequency components.
54
What is quantisation?
1. Take the continuous value.
2. Match it to the nearest available digital value.
3. Store that digital value instead of the original.
-> translates continous to digital
-> allows for as fewer bits as possible
55
What is companding?
Compression + Expanding
Used to avoid unequal levels of noise depending on the amplitude of the signal.
56
What are some companding algortithms?
A-Law (in Europe) and µ-law (US+Japan)
57
What are the sub-processes of analog to digital conversions stages?
- Filtering
- Sampling
- Quantisation
58
What process can you not avoid when you quantize a signal?
* The addition of noise
* Quantization always adds some noise; however, the amplitude of the noise is smaller the larger the bits/sample we choose
59
What is PCM?
- Pulse code modulation is a method used to digitally represent analog signals
- Filter -> compressor -> sampler -> quantizer
Standard PCM channel = 8bit/sample * 8000samples/second = 64Kbps
60
What is the main problem of PCM?
High bitrate
61
What are two different types of compression?
- Compression from companding
- Compression to reduce bitrate of transmission or storage
62
What are simple ways to reduce bitrate?
- sample less often:
1. degrades quality
2. sound may loose its defining characteristics
- sample with fewer bits/sample:
1. large quantisation noise
2. worse SNR (harder to perceive)
63
What are the smarter compressions?
- differential pulse-code modulation
- linear predictive coding
- perceptual coding
64
What are the three differential pulse-code modulations?
1. DPCM
2. Predictive DPCM
3. Adaptive DPCM
65
What is Differential PCM?
- instead of coding the current value, of the signal every time, code the difference
- we can then use fewer bits/sample to code the changes, at a comparable quality.
66
What is predictive differential PCM?
- same concept as DCPM
- encodes difference between predicted sample and current sample
- predicted sample is created several previous samples
67
What is adaptive differential PCM?
- predictive (8 order)
- coefficients change depending on signal
- fewer bits to encode smaller differences
68
What is sub-band coded ADPCM?
A speech/audio compression method that combines sub-band coding with Adaptive Differential Pulse Code Modulation (ADPCM).
How it works:
1. Split the audio signal into frequency bands (sub-bands).
2. Apply ADPCM separately to each band.
3. Encode differences adaptively, reducing bit rate while keeping quality.
Key idea: Human ears are more sensitive to some frequencies than others. By coding each band separately, compression is more efficient and sounds better.
Uses: Broadcast audio, telephony, digital TV, and multimedia systems.
Benefit: Lower data rate with good speech/audio quality.
69
What is linear predictive coding?
Definition: A method to represent speech by predicting each sound sample from past samples.
How it works:
1. Current sample ≈ weighted sum of previous samples.
2. Store predictor coefficients + pitch + loudness.
3. Recreate speech using these values.
Key idea: Models the vocal tract as a filter and the voice as a source.
Uses: Speech compression, synthesis, recognition.
70
What is MPEG?
- Moving Picture Experts Group
- video and audio standards
71
Name several strategies to compress an audio signal.
* Encode change instead of absolute values
* Embed assumptions about the source and
characteristics of the signal
* Split the signal in subbands, and codify subbands differently according to frequency and/or sensibility
* Do not encode what you cannot hear (embed assumptions about human hearing in the encoding of the signal)
72
What is sound?
- variation of pressure
- frequencies in Hz-kHz
- speed ~330m/s
73
What is light?
- variation of em field
- frequencies in 10^14 - 10^15 Hz
- speed ~ 300,000,00 m/s
74
True or false: sound cannot propagate in a vacuum.
True.
75
What are the key parts of the eye?
- pupil
- iris -> providing 20~ fold variation in area
- cornea
- lens
- retina
- optic nerve
- fovea
76
What is photopic vision?
Definition: Daytime vision that works in bright light.
Cells used: Cone cells in the retina.
77
What are cones?
- fovea
- 3 types ( S M L )
- responsible for colour vision and acuity
- about 4.5 mil
78
What is scotopic vision?
Definition: Nighttime vision that works in very low light.
Cells used: Rod cells in the retina.
79
What are rods?
- peripheral vision
- very sensitive
- night vision
- roughly 90 mil
80
What is HDR?
- High dynamic range imaging
81
What is colour perception?
- visible spectrum: 380 - 760 nm
- cones S M and L wavelength sensitivity
- HSV
- Hue: dominant frequency
- Saturation: purity of signal
- Value: brightness intensity
82
What are metamers?
- two different signals can produce the same colour sensation
- means any colour sensation cana be generated
83
what are limits of visual perception
- intensity threshold is very small
- flicker detection -> critical flicker frequency ~120 Hz
- acuity up to 0.5 seconds of arc
84
What are the visual perceptual effects ?
- masking
- blind spot
85
Electromagnetic radiation with wavelengths between 400nm and 700nm is known as?
Visible light
86
True or false: Each cone cell can distinguish a large number of
different frequencies?
False
87
What does a CCD sensor do?
Moves charge pixel-to-pixel, converts to voltage at output
88
What does a CMOS sensor do?
Converts charge to voltage inside each pixel
89
What is the additive colour scheme?
RGB – screens add Red, Green, Blue light to make colours
90
What is the subtractive colour scheme?
CMYK – inks absorb light (Cyan, Magenta, Yellow, Key/Black)
91
What is a colour gamut?
The range of colours a device or ink can reproduce
92
Why do we need colour profiles?
Devices differ – profiles ensure consistent colour reproduction
93
What is colour depth in a 24-bit image?
8 bits per channel (R, G, B)
94
What is resolution?
Number of pixels (e.g., VGA 640×480, 1080p 1920×1080, 4K UHD)
95
What are indexed images?
Use a palette of colours; pixels reference palette entries
96
What is a bitmap image?
Pixel grid, stable, good for photos, not scalable
97
What is a vector image?
Shapes/curves described mathematically, scalable, efficient, good for animation/3D
98
Name one vector format.
SVG, PDF, PostScript, DWG
99
What are basic SVG elements?
Paths, shapes, text, colour, gradients, filters, animation
100
How is SVG a compression mechanism?
Source compression (assumes abstract graphics) + XML lossless compression
101
What are the advantages of vector graphics?
- Scalable (no loss of quality when zooming)
- Efficient for some uses
- Easy to animate
- Good for 3D and transformations
102
What are the disadvantages of vector graphics?
- Render‑dependent
- Encoding varies a lot
- Can store invisible info
- Less realistic
- Lossy compression not possible
103
What are the basic elements of SVG graphics?
- Paths (lines, curves)
- Shapes (circle, rectangle, etc.)
- Text
- Colour (RGB)
- Gradients & patterns
- Clipping, masking, compositing
- Filter effects
- Interactivity (events, focus)
- Linking & scripting
- Animation
- Fonts & metadata
104
What types of noise affect images?
- Quantisation noise → reduce with higher bit depth
- Thermal noise → reduce by cooling
- Background signals → reduce with filtering
- Lossy encoding noise → reduce with lossless or less lossy encoding
105
What is the difference between lossless and lossy compression?
- Lossless: No data lost (e.g. PNG, GIF).
- Lossy: Removes less important info (e.g. JPEG).
106
What is aliasing in images?
- Same problem as undersampling in sound.
- Frequency & sampling are spatial instead of temporal.
- Causes jagged edges or moiré patterns.
- Fixed with anti‑aliasing filters.
107
What are the main steps in JPEG encoding?
What it is: The whole process of turning an image into the JPEG format.
Steps included:
1. Convert RGB → YCbCr (color space conversion).
2. Apply DCT to split the image into frequency components.
3. Quantization (lossy step).
4. Entropy coding (lossless step, e.g., Huffman coding).
Key idea: Encoding = the pipeline that prepares and stores the image as a JPEG file.
108
What are common colour spaces?
RGB → screens, cameras
CMYK → printing
HSV/HSL → perceptual, colour picking
YCbCr → separates luminance (Y) from chromaticities (Cb, Cr); allows reduced resolution for colour
109
How does PNG encoding work?
The whole process of turning an image into the PNG file format.
Steps include:
1. Filtering: rearranges pixel values to highlight patterns.
2. Compression: applies Deflate (LZ77 + Huffman coding).
3. Packaging: stores pixels + metadata (like transparency, color info).
Key idea: Encoding = the pipeline that produces a PNG file.
110
What types of compression are used in JPEG?
Entropy compression (lossless: RLE, Huffman, differential)
Lossy perceptual compression (quantization, undersampling colour components)
111
True or false: Macroblocks for luminance carry more information than the macroblocks for chromaticity
True
112
Which factors influence the level of
compression?
A. Content complexity
B. Settings of compression algorithm
D. Implementation of the encoder
113
Do macroblocks for luminance carry more information than macroblocks for chromaticity?
Yes – luminance macroblocks (16×16) carry more information than chromaticity (2×8×8).
114
Which factors influence the level of compression?
Content complexity & movement, compression algorithm settings, and encoder implementation.
115
What are some common tasks in signal/image analysis?
Peak finding, mean intensity (over time), and image segmentation.
116
What is image segmentation?
Dividing an image into meaningful regions (e.g., organs, tumours) using edge detection, local features, or machine learning.
117
What measurements are taken when characterising nanorods?
Length, width, aspect ratio (L/W), and histograms of these values.
118
What is the watershed algorithm used for?
Segmenting images by treating intensity as topography, identifying catchment basins and watershed lines.
119
What exclusion criteria are used for detected objects?
Area too small/large, eccentricity <0.7, too long/short, or touching image border.
120
What imaging modalities are used for brain tumour segmentation?
MRI (T1, T1c, T2, FLAIR), CT, and PET.
121
What datasets are widely used for brain tumour segmentation research?
BraTS (Brain Tumour Segmentation Challenge) dataset.
122
What progress has CNNs brought to image segmentation?
Significant improvements in semantic segmentation, especially for multi-class tumour sub-regions.
123
What metric is commonly used to evaluate segmentation quality?
Dice coefficient: =
2𝑇𝑃
---------------------------
2𝑇𝑃+𝐹𝑃+𝐹𝑁.
124
What is pathology? Histopathology?
1. Pathology is the science of diseases. It studies their causes and effects by looking at tissue samples.
2. Histopathology is the study of changes in tissues caused by disease.
125
What is digital pathology?
Digital pathology uses scanners to turn tissue slides into digital images. These can be stored, shared, and analysed by computers.
126
What are foundation models in pathology?
Large models trained on millions of slides (e.g. UNI, Phikon, CTransPath). They learn general features and can be fine‑tuned for tasks.
127
What is a convolution?
A way to combine an image with a filter (small matrix) to change or highlight features.
128
What can convolutions be used for?
To filter images, detect edges, blur, sharpen, or segment parts of an image.
129
What is a digital image?
A grid of numbers (pixels). Each number shows brightness or colour.
130
What is linear filtering?
Applying a filter (matrix) to an image by multiplying and adding values.
131
What is the difference between correlation and convolution?
1. Correlation: filter slides over image without flipping.
2. Convolution: filter is flipped before sliding.
132
Why do we use padding?
To handle edges of an image when applying filters.
133
Types of padding?
1. Zero padding: add zeros around edges.
2. Constant padding: add fixed values.
3. Replicate padding: copy edge values.
4. Reflect padding: mirror the image edges.
134
What is Otsu’s method?
A way to automatically find the best threshold to separate foreground and background.
135
What is the watershed algorithm?
A method that treats the image like a landscape and finds “basins” to separate regions.
136
Why are convolutions important in computer vision?
They help machines see patterns in images, like edges, shapes, or textures.
137
What is a common use of convolutions in AI?
Convolutions are the building blocks of Convolutional Neural Networks (CNNs).
138
What does machine learning mean?
It means a computer learns from data instead of being given fixed rules.
139
Why is Machine Learning useful?
It helps us find patterns, make predictions, and handle complex tasks that are hard to code by hand.
140
What does a ML model do?
It takes input data, learns a pattern, and gives a prediction.
141
What are the two main types of machine learning?
Supervised learning and unsupervised learning.
142
What is supervised learning?
Learning from labelled data where we know the correct answers.
143
What tasks use supervised learning?
1. Predict numbers (regression)
2. Predict categories (classification)
144
What is unsupervised learning?
Finding groups or patterns in data without labels.
145
What is linear regression?
To predict a number using a straight‑line relationship.
146
What is overfitting?
When a model learns the training data too well, but fails on new data.
147
What is underfitting?
When a model is too simple and misses important patterns.
148
How do we avoid overfitting?
Split data into training, validation, and test sets.
149
What is classification?
Predicting a class label, like “cancer” or “not cancer”.
150
What is a decision boundary?
A line or shape that separates classes in the data.
151
What is a neural network?
A model inspired by the brain, made of layers of neurons that learn patterns.
152
What is a sigmoid function?
It turns numbers into a value between 0 and 1, useful for classification.
153
What is “one‑vs‑all” classification?
Train one model per class.
Each model says: “This class vs. all others.”
Example: Cat vs Not‑Cat, Dog vs Not‑Dog, Bird vs Not‑Bird.
154
Where is machine learning used?
Self‑driving cars
Medical imaging
Banking
Online shopping
Voice assistants
155
What is a feature?
A piece of input data, like age, pixel value, or test result.
156
What is a hyper‑dimensional boundary?
A decision boundary in many features, too many to draw.
157
What are common ML algorithms?
Logistic regression
SVM
Neural networks
Random forests
158
What is the k-means algorithm used for?
It groups data into clusters by moving centroids to the average position of points.
159
What is a neural network inspired by?
The human brain - neurons connected by input and output wires.
160
What is max pooling?
Idea: downsampling method - reduce the size of feature maps to make the network faster and more focused.
What it does: Takes the largest value from a small block (e.g. 2×2) - then summarises that block into one single value
Why: Keeps the strongest feature (like sharp edges).
Effect: Shrinks data size, highlights important details.
161
What is average pooling?
Idea: downsampling method - reduce the size of feature maps to make the network faster and more focused.
What it does: Takes the average value from a small block - then summarises that block into one single value
Why: Smooths features, keeps overall trend.
Effect: Shrinks data size, reduces noise, less sharp than max pooling.
162
What does gradient descent do?
It finds the best parameters by moving step by step to reduce error.
163
What is a confusion matrix?
A table showing true positives, true negatives, false positives, false negatives.
164
What are X‑rays best for?
Imaging bones and detecting fractures.
But can also be used for detecting pneumonia and for mammography.
165
What does ultrasound use instead of radiation?
High‑frequency sound waves (1–15 MHz).
166
What are the main ultrasound modes?
1. A‑mode (Amplitude mode)
Shows echoes as spikes on a graph.
Used for measuring distances or tissue thickness (e.g., eye exams).
2. B‑mode (Brightness mode)
Produces a 2D grayscale image.
Most common mode for viewing organs and tissues.
3. M‑mode (Motion mode)
Tracks movement over time (one scan line).
Often used for heart motion and valve studies.
4. Doppler mode
Measures frequency shifts in echoes.
Shows blood flow speed and direction (color Doppler adds visual maps).
167
Why are ultrasounds safer than CT scans?
Because it doesn't use ionising radiation.
168
What do MRI scanners align in the body?
Hydrogen nuclei (protons) in water.
169
What is TR (Repetition Time) and TE (Time to Echo) in MRIs?
TR: Time between successive RF pulse sequences.
TE: Time between RF pulse and echo signal.
170
How do T1 and T2 images differ in MRIs?
T1: Fat bright, CSF dark
T2: Fluid bright, CSF bright
171
What does FLAIR imaging do?
Makes CSF dark but keeps abnormalities bright → easier to spot pathology.
172
What does PET measure? What is PET useful for?
Measure: Metabolism (sugar uptake in tissues).
Use: Detecting cancer spread (metastasis).
173
What is endoscopy?
Using a camera inside the body to see organs directly.
174
What happens when a positron meets an electron in PET scans?
They annihilate → two gamma rays (511 keV) emitted in opposite directions.
175
What does a CT scanner use? What planes are scanned the most?
Rotating X‑ray source and detectors with a moving table.
Planes: Coronal (front/back), Sagittal (side), Axial (top/bottom).
176
Why can X‑rays be harmful?
They can damage DNA through ionisation.
177
What is back projection? Filtered back projection?
Back: Spread beam values across pixels, add them up to estimate the image.
Filtered: Same idea, but uses filters to reduce blur.
178
What are Hounsfield units in CT?
A scale for CT values (air ≈ ‑1024, water ≈ 0, bone ≈ +1024).
179
Why is registration needed in imaging? How can registration be done?
1. Patients move, scans differ in angle or slice thickness.
2. Use markers (internal or external) that appear in all images.
180
What types of noise affect images and how can noise be reduced?
Electronic, thermal, speckle (ultrasound), artefacts.
Reduced: Filters, subtract dark current, average over time.
181
What are ROIs?
Regions of interest.
182
How can ROIs be found?
Manual boxes, image analysis, ML
183
What does a pulse oximeter measure and what levels are considered dangerous?
- heart rate: HR < 40, > 130 bpm
- oxygen: <90%
184
What is video encoding? Why do we need it?
What: Turning video into a digital format that can be stored or sent.
Why: Raw video is huge. Compression makes files smaller by removing repeated or less important data.
185
How many images per second can the brain process?
Less than about 20 per second. The brain holds each image briefly, hence why the fast updates feel continuous and look smooth.
186
How does a CRT show video?
It scans lines across the screen with light.
187
What is interlacing?
Splitting even and odd lines to reduce flicker.
188
Why is uncompressed digital video expensive?
Example: 1080p at 30 fps needs ~1.5 billion bits per second.
-> need to reduce and remove redundancy
189
What is chroma subsampling (compression trick)?
Store less colour detail than brightness detail (e.g., 4:2:0).
190
What is frequency quantization (compression trick)?
Keep low‑frequency detail, drop high‑frequency detail (like JPEG).
191
What is temporal redundancy (compression trick)?
Frames are similar. Only send differences.
192
Typical compression ratios?
I‑frame: 10–20:1
P‑frame: 20–30:1
B‑frame: 30–50:1
193
What is the goal of a data science pipeline?
To move from a question to clear results through steps like getting data, exploring it, modelling it, and reporting findings.
194
What do you do when exploring data?
Split data, check for missing values, look at class balance, view samples, and calculate simple stats.
195
What is the conventional image analysis pipeline?
Acquire → clean → label → handcrafted features → background removal → edge detection → segmentation → rule‑based classification.
196
What is the machine learning image pipeline?
Acquire → clean → label → split data → extract features → train ML model → evaluate.
197
What is the deep learning image pipeline?
Acquire → clean → label → split data → build CNN → train → evaluate.
198
What are key regression metrics?
MAE, RMSE, and R².
199
What are TP, TN, FP, and FN?
TP: correct positive.
TN: correct negative.
FP: predicted positive but wrong. FN: predicted negative but wrong.
200
What are key classification metrics?
Precision, recall, F1 score, accuracy, and confusion matrices.
201
What is optical flow? Farneback optical flow? Lucas–Kanade optical flow?
1. Measuring motion between video frames.
2. Calculates motion vectors for every pixel.
3. Calculates motion for selected points or regions.
202
What are artefacts in images/videos?
Things that look different from reality. They can come from capture, encoding, or playback.
203
What is thermal noise?
Random electrical noise in sensors → lowers signal quality.
204
What is quantisation noise?
Small errors when turning analogue signals into digital numbers.
205
Why does motion affect compression?
More motion → harder to compress → lower quality at fixed bitrate.
206
What if frame rates differ?
Frames are duplicated or interpolated → may look odd.
207
What happens when image size changes?
Scaling up/down can cause blur or blockiness.
208
What is colour mapping?
Converting stored colours to screen RGB. Needs careful scaling if >8 bits per channel.
209
What is ImageNet?
- A dataset with 14M+ images, 1M+ bounding boxes, 20k+ categories.
- Allowed for breakthroughs like AlexNet and modern deep learning.
- Deep conv nets allowed for over 95% accuracy on classification tasks.
210
What is the Cancer Imaging Archive?
120+ open datasets (MRI, CT, PET, Ultrasound) for cancers.
211
What is CheXpert?
A huge X‑ray dataset from Stanford.
212
What are the two compression methods for video?
1. Intra-frame: Spatial redundancy, compress each frame like JPEG
(to remember: r comes first in intra and r is next to s in the alphabet hence spatial)
2. Inter-frame: Temporal redundancy, motion compensation, predicative coding.
3. (t come after s in the alphabet so after r is inter, and hence temporal)
213
What is the video classification pipeline?
1. set-up (lighting, etc)
2. audio (spectrograms, species specific ranges, etc)
3. video (CNNs for visual classification)
4. combine video and audio
214
How does a CT scanner work?
X‑ray tube rotates around patient.
Detectors measure how much X‑ray passes through.
Computer combines many angles → cross‑section image.
Key idea: Shows tissue density differences.
Lossy step: X‑ray dose absorbed by body.
215
How does MRI work?
Strong magnet aligns hydrogen atoms.
Radio waves knock them out of alignment.
As they relax, they send signals.
Gradients map position → image.
Key idea: Great soft tissue contrast.
216
How does PET work?
Inject radioactive tracer (e.g. FDG).
Tracer emits positrons → annihilation → 2 gamma rays.
Detectors in ring catch both rays.
Computer maps tracer distribution.
Key idea: Shows metabolism, not structure.
217
How does ultrasound work?
Probe sends sound waves into body.
Echoes return from tissue boundaries.
Time + strength of echoes → depth + brightness.
Computer builds real‑time image.
Key idea: Safe, portable, live imaging.
revision
flashcards
Decks in class (12)
# Cards
