Biomedical Signal and Image Processing
This domain tests whether you can choose and justify processing steps for physiological signals and medical images, and evaluate the result. Questions combine filter design, sampling arithmetic, image formation and quantitative performance metrics.
Signal processing essentials
- Sampling theorem: fs > 2 fmax, with an analogue anti-alias filter before the ADC. In Hong Kong the mains component is 50 Hz at 220 V, so a notch at 50 Hz plus harmonics is standard for ECG and EEG.
- FIR filters give exactly linear phase and guaranteed stability at the cost of longer kernels; IIR filters achieve sharp cut-offs with fewer coefficients but have non-linear phase and stability risk.
- ECG chain: bandpass 0.05–150 Hz, notch, baseline-wander removal (high-pass or morphological), then QRS detection (Pan–Tompkins uses derivative, squaring, moving-window integration and adaptive thresholds).
- Adaptive filters (LMS) are used when the noise reference is measurable, for example motion artefact for wearables.
Imaging
- CT reconstruction uses filtered back projection or iterative methods; Hounsfield units HU = 1000 × (µ_tissue − µ_water) / µ_water, with water 0 HU and air about −1000 HU.
- Windowing maps a HU range to display grey levels: window width is contrast, window level is brightness.
- MRI encodes spatial position in k-space; undersampling accelerates acquisition but requires reconstruction that handles aliasing.
- DICOM defines image objects and network services; PACS stores and distributes them.
Worked example: aliasing
A monitor samples an ECG at 100 Hz. A 60 Hz interference component folds to |60 − 100| = 40 Hz, landing inside the diagnostic band and masquerading as a physiological rhythm. Increasing fs to 500 Hz moves the alias to 440 Hz, outside the band, and an analogue low-pass at 150 Hz removes it entirely. State both fixes.
Worked example: CT windowing
A soft-tissue window is set to level 40 HU and width 400 HU, so the display spans −160 to +240 HU. Fat near −100 HU appears dark grey, while bone above 240 HU saturates white.
Evaluating performance
- Sensitivity and specificity describe the test; PPV and NPV depend on prevalence, so a 90% sensitive, 95% specific test at 1% prevalence gives PPV ≈ 0.9×0.01 / (0.9×0.01 + 0.05×0.99) ≈ 15%.
Traps that cost marks
- Quoting Nyquist without an anti-alias filter or ignoring the ADC's own bandwidth limits.
- Confusing window width with window level.
- Reporting only sensitivity when a screening decision needs predictive values.
- Using IIR filters where phase distortion would misplace a fiducial point.