Medical sensing, imaging and diagnostic-system principles

A physical signal needs an interpretation

A sensor responds to a physical, chemical or biological input. Its signal passes through amplification, filtering, sampling, calibration and an interpretation model. State the measurand and where it is sensed. A quantity measured in one fluid or at one body location need not equal a different quantity elsewhere without a justified relationship.

Motion, contact, interfering substances, drift and timing can affect the signal or the mapping to a reported result. A device may return precise repeated outputs while having a systematic error. Preserve units, processing version, measurement conditions and relevant uncertainty. In biomedical use, the meaning of an output depends on the whole measurement chain and population.

Different imaging methods encode different interactions

Computed tomography uses x-ray measurements from multiple directions and reconstruction to form cross-sectional images. MRI uses magnetic fields, radiofrequency excitation and detected nuclear magnetic signals to encode spatial information. CT involves ionizing x-rays; MRI does not use ionizing radiation, though its fields and equipment still require specific safety controls.

Resolution, contrast, noise, motion and reconstruction affect what can be distinguished. A small displayed pixel does not guarantee an equally small feature is resolved. Image appearance and an automated label also depend on acquisition and processing. Choosing or interpreting an examination belongs to qualified clinical work; here the engineering task is to understand what information the instrument supplies.

A ray mixes locations; several views constrain them

In CT, some x-ray photons are removed from a narrow transmitted beam by interactions along its path. A detector receives the remaining beam. One ray therefore contains information from several locations, rather than a label for one internal point. Reconstruction uses known ray geometry and many views to infer a spatial attenuation map. As in the instrumentation preparation, a detector signal and an inferred measurand are different stages.

Use a deliberately tiny two-by-two model of uniform square cells, with nonnegative dimensionless attenuation contributions a and b in the top row and c and d below. Each number describes a one-cell-side path. Supply already converted line-attenuation readings: row sums a+b and c+d, and column sums a+c and b+d. These are not raw photon counts: the transmitted fractions multiply through successive cells, and the conversion to attenuation is what makes this ideal path account additive. Scatter, a spectrum of x-ray energies, noise and motion are omitted.

Map A has rows [1,3] and [3,1]; map B has [3,1] and [1,3]. Both give row readings 4 and 4 and column readings 4 and 4. The four readings do not locate the larger contributions uniquely. A diagonal central ray through a and d travels √2 times the cell-side distance in each. Divide its converted reading by that known factor to obtain a+d: 2 for A and 6 for B. This additional view distinguishes the two maps. More displayed pixels could not have supplied that missing measurement.

This is a small inverse problem: seek an internal map consistent with external measurements and explicit assumptions. Real reconstruction needs much richer sampling and uncertainty treatment. If the object moves between views, they may no longer describe one common map. A smooth or sharp-looking reconstruction is not by itself evidence that the information needed for a particular feature was acquired.

MRI needs a spatial label as well as a signal

In the usual hydrogen MRI picture, nuclear magnetic moments in a strong static field produce a small net magnetization. An applied radiofrequency field can tip that net magnetization away from its equilibrium direction. After excitation, part of the magnetization rotates about the static-field direction; this motion is called precession. The resulting changing magnetic signal induces a voltage in a receiving coil. The measured signal depends on the amount and state of magnetization, its evolution and the receiving system. It is not an x-ray transmission count or an immediate photograph of individual protons.

A coil can receive a mixture from an extended region. Controlled spatial variations in the magnetic field, called gradients, make precession frequency depend on position. In a simple one-dimensional readout, frequency can then label location along the gradient. Other encoding steps distinguish the remaining directions; their timing also records phase, the position within an oscillation cycle. The computer reconstructs an image from these combined signals. Relaxation, meaning return toward equilibrium and loss of coordinated rotation, affects the signal. Receive sensitivity, motion and incomplete encoding also affect what can be inferred; signal amplitude alone is not a universal tissue label.

For a wholly supplied frequency-to-position illustration, let x run from 0 to 2 arbitrary length units and let the frequency offset from a common carrier be 100x Hz. A component at 50 Hz is associated with x=0.5, and one at 150 Hz with x=1.5, provided no other frequency shift is present and the encoding relationship is known. Without the position-dependent encoding, two regions with the same resonance frequency would contribute to one mixed signal without that location label. The plotted offsets are not field settings, an acquisition sequence or a resolution guarantee.

A constructed diagnostic-result model

Assume a hypothetical binary test has sensitivity 90% and specificity 95% in a specified population, with a perfect reference classification and one outcome per person. In an invented group of 1000, 100 have the target condition. Expected counts are 90 true positives, 10 false negatives, 45 false positives and 855 true negatives. These are model counts, not observations from a device study.

Sensitivity concerns positives among those with the condition: 90/100. Specificity concerns negatives among those without it: 855/900. Among positive results, the condition is present in 90/(90+45)=66.7%. The figure preserves all four cells so that the denominator of each measure remains visible. “90% sensitive” therefore does not mean that 90% of positives have the condition.

Change prevalence while holding the model’s test rates fixed

In a second invented group of 1000, let only 20 have the condition. Under the same supplied conditional rates, expected counts are 18 true positives, 2 false negatives, 49 false positives and 931 true negatives. Positive predictive value becomes 18/67≈26.9%, even though sensitivity and specificity were held fixed by assumption.

Actual estimated rates have uncertainty and can change with population spectrum, methods and reference quality. Agreement with another imperfect test is not automatically correctness. Invalid or indeterminate outcomes and exclusions must remain visible. The example explains conditional probabilities; it does not recommend screening, diagnose an individual or establish a real test’s performance.

For the CT toy model, suppose all four row/column readings are 4 and the normalized diagonal reading a+d is 2. The row and column equations give d=a and b=c=4−a. Thus 2a=2, so a=d=1 and b=c=3: map A. The additional view resolves this particular ambiguity under exact, stationary assumptions. It does not prove that two orthogonal views plus one ray reconstruct an arbitrary real object.

Try the reasoning

1. Why is positive predictive value different from sensitivity?

   Show the answer and reasoning

   The first conditions on a positive result; the second conditions on the target condition being present. Their denominators differ.

2. Does a small image pixel guarantee that a feature of that size is resolved?

   Show the answer and reasoning

   No. The acquisition, system resolution, contrast, noise and reconstruction also limit distinguishable features.

3. Can high agreement with another test alone establish diagnostic accuracy?

   Show the answer and reasoning

   No. Both can agree while being wrong; the benchmark and study design determine what performance can be inferred.

4. The CT row/column readings stay at 4, but the normalized a+d reading is now 6. Which map fits? In the separate MRI illustration, what position corresponds to 150 Hz if the known encoding is reversed to 200−100x Hz?

   Show the answer and reasoning

   For CT, d=a and 2a=6, so a=d=3 and b=c=1: map B. Both maps still fit the original row/column data; the changed view supplies the distinction. For MRI, 150=200−100x gives x=0.5 rather than the original 1.5. The frequency needs its actual encoding relationship to become a position. Neither calculation establishes diagnostic accuracy or achievable spatial resolution.

What to carry forward

Keep the physical signal, reconstruction or inference, and population-dependent performance measures explicit throughout the measurement chain.