How a computer vision model processes a head CT
When someone arrives with signs of a stroke, a CT scan of the head is usually one of the first images taken, partly to check for bleeding in the brain. To a trained reader that scan carries clinical meaning. A computer vision model receives numbers. This page follows the path from stored values to displayed image to model output, using a synthetic image you can zoom into, change, and test.
No math required. The image and the scoring system here are both invented for teaching, and no real patient is involved.
What you will be able to explain
- How a displayed image becomes numerical input for a computer vision model.
- How classification, detection, segmentation, and response maps differ.
- Why a model score, a probability, and a clinical diagnosis have different meanings.
- Which evidence and monitoring questions belong in a real clinical AI review.
Choose the depth you need
Essentials, about 12 minutes. Follow the five links below, then complete the quick check.
Full walkthrough, about 20 minutes. Work through every section and try each control.
What a person sees
A trained reader looks at this and sees anatomy: skull, brain, the fluid-filled spaces in the middle, and possibly something that stands out from its surroundings. Meaning arrives all at once.
Synthetic teaching image · not a real CT scan
Drawn by code from a fixed random seed. Every learner sees exactly the same image.
Two different acts of seeing
A person sees a brain, and inside it maybe a bleed. Those are human concepts, built from years of training and from knowing what a body is. A reader also knows the convention: the patient's right side appears on the left of the image, as if you stood at the patient's feet looking toward their head.
A model receives numerical inputs. During training, it can learn internal patterns associated with labels such as "brain" or "hemorrhage." Those learned associations differ from human clinical understanding and can fail when a new image differs from the training data.
On this page, the input is a rectangle of numbers describing the brightness of each tiny square.
How this differs from a real scan
- A head CT is a stack of thin slices, from the base of the skull to the top of the head. Readers scroll through all of them, often in more than one display window. This page shows one slice.
- Stroke imaging often adds CT angiography, which uses contrast dye to show the arteries. Some of the best-known stroke AI tools analyze those images for a blocked large artery and alert a specialist.
- Most strokes come from a blocked artery. In the first hours, that kind of stroke can look normal or nearly normal on a plain CT. This page draws a bleed because blood shows up as a bright spot you can find at the pixel level.
- This page uses one simplified synthetic slice so that the pixel-level journey stays visible. Treat it as a diagram.
Zoom in far enough and the picture disappears
A CT scanner stores a number for every tiny square that converts to a density measure called the Hounsfield unit (HU). A display window picks which range of those numbers fills the scale from black to white. Zoom in, or click the image, and a teal box marks the area being magnified. At full magnification each square of brightness is a single number. Whichever window you choose here, the simulation always receives the brain window.
Click or tap the image to move the box. With the box focused, use the arrow keys; hold Shift for larger steps.
Three steps: whole image, a 32 by 32 patch, then an 8 by 8 patch with the numbers shown.
0 to 80 HU spread from black to white. Brain, fluid, and blood separate clearly. Bone is pure white.
Magnified region
A grid of 64 pixels. Use the arrow keys to move between them, Enter or Space to pin one. Each number is that pixel's brightness in the chosen window, from 0, black, to 255, white.
Stored value, displayed brightness, and model input
Typical Hounsfield unit (HU) values: air sits near -1,000, water at 0, brain between about 20 and 45, a fresh brain bleed around 50 to 65, and bone in the hundreds or more. No screen can show that whole range as useful shades of gray at once.
So the reader chooses a window. A brain window spreads 0 to 80 HU from black to white; anything denser shows as pure white. A bone window spreads a far wider range, which reveals the skull and flattens the brain. Switch between them above: the stored numbers stay the same while the picture changes.
This page simplifies two things:
- The simulation receives one brain-windowed image with 256 shades. Real pipelines choose their own windows and value ranges.
- Model pipelines commonly preprocess images by cropping, resizing, rescaling, or filtering them. The order and exact steps vary by system.
From picture to array
Here is the patch you selected, written out as numbers. These numbers always come from the brain window, the version the simulation receives, whatever window you chose above. In this simulation, a grid like this repeated across the whole image is the complete input. Real systems may also use every slice of the scan, prior scans, or metadata.
The stored values, converted to Hounsfield units.
The whole journey, in four steps
Resizing in this simulation
Full image
256 × 256
Model input
64 × 64 = 4,096 numbers
The simulation on this page shrinks the image to 64 by 64 before analyzing it, which is why the version on the right is the one that gets scored. Real systems use larger inputs, and choosing that size is a design trade-off between fine detail and computation.
How this teaching simulation builds a score
This page applies fixed comparisons written by its author. The sequence offers simplified intuition for one convolution-style workflow: local differences become edges, focal responses, coarse regions, and finally a score. Before any of that, it strips away the skull. Bone stores far higher values than brain or blood, so this simulation treats the brightest pixels as bone and keeps only the brain. On screen, a brain window shows everything above 80 HU as pure white, so very dense blood can look as white as bone, and a bleed pressed against the skull can blend into it. Some published and FDA-authorized head CT pipelines include a skull-stripping step, and a real pipeline needs checking for bleeds lost at the brain's edge. Step through the teaching simulation below.
Simulation step 1
Pixel differences
Is this number bigger or smaller than its neighbor? Nothing more.
Simulation step 2
Edges and textures
Differences that line up become boundaries and grain.
Simulation step 3
Shapes and densities
Compact areas that stand out from what surrounds them.
Simulation step 4
Coarse regions of response
The strongest response in each region is recorded for this display.
Simulation step 5 · a single number comes out
The simulation combines four measured features: the strongest focal response, how concentrated that response is, how sharp it is, and how much grain is present. Fixed rule-based calculations turn those measurements into one output. A trained model learns its weights from examples, which is the subject of the next lesson.
Classification, detection, segmentation
Classification, detection, and segmentation answer different technical questions. The fourth tab shows the internal filter response from this teaching simulation. A predictive localization map, an internal response map, and a post hoc attribution map each represent a different quantity.
Simulated output · synthetic image
Classification
| Concept | Meaning | What comes out |
|---|---|---|
| Image classification | Assigns categories or scores to a defined unit, such as a slice or a whole scan | One or more labels or scores |
| Object detection | Identifies candidate locations of possible findings | Zero or more boxes, points, or markers, often with per-candidate scores |
| Segmentation | Estimates the boundaries of a finding | Per-pixel scores that are commonly converted into an outline or mask |
| Predictive localization map | Produces location-specific scores as a trained model output | A map of location-specific or per-pixel scores |
| Internal response map | Shows a defined internal quantity, such as this simulation's filter response | A shaded overlay tied to a specific calculation |
| Attribution or saliency map | Estimates which parts of an input influenced a particular output | An explanatory overlay that requires method-specific validation |
A score, a probability, and a diagnosis have different meanings
Everything so far produces one number. The model's documentation and validation determine what that number means.
Simulated suspicion score
…
…
A threshold turns a continuous score into a yes or no. Move it and watch the same score cross the flagging threshold. In the FDA-authorized triage device cited below, the manufacturer set and tested a fixed threshold.
What this number is
- The output of this simulation on this synthetic image, produced by fixed filters written for teaching.
- A relative measure. Higher means the image contains more of what these filters respond to.
- A quantity that a threshold can convert into an action, such as moving a scan up the reading list or alerting a specialist or the stroke team.
Limits of this number
- A bleed can be present after a low score and absent after a high score.
- A score of 0.72 remains a unitless score. It represents a 72% estimated probability only when the model has been designed, evaluated, and shown to support that interpretation in the intended population.
- The radiologist and the surrounding clinical workflow determine what happens next.
Four terms that make threshold trade-offs precise
- Sensitivity is the share of cases with the condition that the system flags.
- Specificity is the share of cases without the condition that the system leaves unflagged.
- Positive predictive value is the share of flagged cases that have the condition.
- Negative predictive value is the share of unflagged cases that do not have the condition.
The threshold changes sensitivity and specificity. Predictive values also depend on how common the condition is in the population being scanned. Every metric needs a defined population, reference standard, threshold, and measure of uncertainty.
Watch one threshold affect a group
This teaching cohort contains 100 invented scans: 20 where a bleed is present and 80 where it is absent. The scores demonstrate the math and represent no clinical system, real rate of bleeding, or expected performance.
| Reference standard | Flagged | Below threshold | Total |
|---|---|---|---|
| Condition present | … true positives | … false negatives | 20 |
| Condition absent | … false positives | … true negatives | 80 |
Image conditions change what the model receives
The first five controls change the image conditions while the drawn finding stays fixed. The final control changes how clearly that finding appears against its surroundings. Some changes barely move the score. Others move it a lot. That gap is one of the most important things to understand about computer vision.
Teaching simulation · not a validated model
Score now
…
Change from baseline
…
Adjustments
A schematic global shift, similar to moving the window level. Real scanner and protocol differences affect images in more complex ways.
A schematic grain control. Real CT noise depends on radiation dose, patient size, reconstruction settings, and the scanner.
A schematic proxy. A tilted head on a real scanner changes which structures each slice cuts through.
Stands in for a streak artifact or an area the scan did not capture. The band covers from the right edge.
Changes how dense the drawn bleed is compared with the brain around it. This control is included for comparison with the image-condition changes above.
Why this matters outside a teaching page
A model is fitted to the images it was shown. When the images it meets later differ, performance can differ too, even when the anatomy is identical. Common sources of that difference include:
- Different scanner makes and models
- Image quality and dose
- Head position and tilt
- Preprocessing choices
- Motion, metal, and other artifacts
- Missing slices or incomplete coverage
- Differences between hospitals and populations
- Shortcuts tied to labels, markers, borders, scanners, sites, or processing
Watching this over time, deciding when a difference matters, and knowing what to do about it are their own subjects. Later lessons cover monitoring and drift.
The seven things worth carrying away
- Triage, diagnosis, and treatment serve different purposes. Triage software flags a scan so it is read sooner or sends an alert to a specialist. The radiologist still reads every scan. Treatment decisions draw on that read, the patient's examination, and the wider clinical picture.
- A flag does not mean a bleed is present. Every threshold trades missed cases against false alarms, and each false alarm costs a team time and attention.
- An unflagged scan does not rule out a bleed or a stroke. Findings can be missed, by people and by models. About 87% of strokes come from a blocked artery, and those can be hard to see on a plain CT in the first hours. Some tools also skip scans they cannot process, such as incomplete studies or scans with metal clips. A skipped scan gets no flag, so check how the tool reports the scans it did not analyze.
- A response map needs a precise name and evidence. Saliency, activation, attention, lesion-likelihood, and filter-response maps represent different quantities. A convincing overlay can still give a poor account of model behavior.
- The score on this page is illustrative. It came from rule-based filters over a drawn image.
- Evidence must match the intended use. Look for representative data, independent and multisite testing, relevant subgroup results, local workflow evaluation, prospective evaluation when feasible, applicable regulatory review, planned human-device interaction, and ongoing monitoring.
- Nothing here supports a medical decision. If you think someone is having a stroke, call 911 or your local emergency number right away.
Evidence before use
- What is the intended use, input, output, user, and place in the workflow?
- How was the reference standard, the method used to determine the correct answer, established? Who assigned labels, what follow-up was available, and how were uncertain or disputed cases handled?
- Were training, tuning, and test partitions separate at the patient level or higher? What controls kept information from test cases out of model development, a problem called data leakage?
- Was the completed model tested on external or otherwise development-independent data spanning representative sites, scanners, protocols, populations, and subgroups relevant to the planned use?
- Are statistical uncertainty and subgroup results reported alongside overall performance? When outputs are interpreted as probabilities, is calibration reported?
- How does the human-AI team perform compared with the current clinical workflow?
Controls after launch
- Who verifies inputs and outputs, who can override the system, and where are disagreements or unexpected results reported?
- Who owns monitoring, which metrics and subgroups are reviewed, and how often does review occur?
- How will changes in equipment, software, workflow, or patient population trigger reassessment?
- What thresholds trigger investigation, temporary pause, revalidation, or retirement, and who has authority to act?
Regulatory note. As of September 29, 2026, the FDA lifecycle source linked below is still the January 2025 draft guidance. FDA labels it "Not for Implementation" and states that it contains nonbinding recommendations.
Keep the reader's own skill sharp
A stroke team that trusts a triage alert still needs radiologists who can spot a bleed on their own, including on the night the software is down. When a tool flags findings every day, people can start to lean on it, and a skill that goes unused can fade. This is called deskilling.
The clearest early signal comes from colonoscopy. In one multicenter observational study at four centers in Poland, 19 experienced endoscopists found precancerous growths (adenomas) in 28.4% of the colonoscopies they did without AI in the three months before AI came into routine use, and in 22.4% in the three months after. A larger single-center study found that detection without AI held steady. Both studies are observational, so the question is still open, and the risk is worth planning for.
No method has yet been proven to prevent AI-related deskilling. Evidence from related settings points to two habits:
- Practice without the tool on purpose. Use training cases or simulation where readers commit to their own call before they see the AI's, then get feedback. In a randomized trial, residents retained more after repeated testing with feedback than after restudying the same material. In a nonclinical experiment, designs that made people answer first, wait for the AI's advice, or ask for it reduced overreliance on the AI, and people rated the designs that helped most the least favorably. In patient care, follow the approved workflow.
- Measure detection without the tool over time. Both colonoscopy studies measured detection in procedures done without AI after AI arrived. A team can track the same thing for its own readers, so a decline shows up early.
A vendor shows a score and a heatmap
A vendor says one head CT received a score of 0.78 and the heatmap shows where the bleed is. Which response best applies the ideas from this lesson?
Choose a response to see feedback.
Sources and technical notes
- Prabhakaran S, Gonzalez NR, Zachrison KS, et al. 2026 Guideline for the Early Management of Patients With Acute Ischemic Stroke: A Guideline From the American Heart Association/American Stroke Association. Stroke. 2026;57(8):e316–e436. Erratum in Stroke. 2026;57(8):e461–e467. Replaces the 2018 guideline and its 2019 update. Section 3.2 states that brain imaging is essential to exclude hemorrhage before clot-dissolving or clot-removal treatment, that CT-based protocols are more efficient and generalizable, and that imaging of the arteries to find a large vessel occlusion should be performed rapidly when clot removal is being considered. Supports the opening of this page and the CT angiography note in section 1.
- Radiopaedia. Windowing (CT). Gives the brain window as width 80 and level 40, so values above +80 HU show white and values below 0 show black, and a bone window of width 1800 and level 400. Supports the window settings used in sections 2 and 3.
- Radiopaedia. Intracerebral haemorrhage. Describes acute bleeding on CT as a bright (hyperdense) collection, often with a darker rim of swelling. Supports how the synthetic bleed is drawn.
- DICOM PS3.3, CT Image Module (C.8.2.1). Defines Rescale Slope and Rescale Intercept, which convert each stored value to output units (output = m × stored value + b), and states that for original CT images the output units are Hounsfield units. Supports the Hounsfield unit statements in sections 2 and 3.
- DICOM PS3.3, VOI LUT Module (C.11.2). Defines Window Center and Window Width, the settings behind a display window. Supports the windowing explanation in section 2.
- DICOM PS3.3, Image Pixel Module (C.7.6.3). Defines how stored pixel samples are encoded (Pixel Data, Bits Allocated, Bits Stored, Photometric Interpretation), including arrangements beyond the 8-bit values used by this teaching page. Supports the section 2 point that stored values differ from the 8-bit display values used here.
- Radiopaedia. Hounsfield unit. Lists typical values: air -1,000 HU, water 0 HU, gray matter 40 HU, white matter 30 HU, trabecular bone 300 to 800 HU, and cortical bone above 1,000 HU. Supports the tissue values in sections 2 and 3.
- Jeong HG, Bang JS, Kim BJ, Bae HJ, Han MK. Hematoma Hounsfield units and expansion of intracerebral hemorrhage: a potential marker of hemostatic clot contraction. International Journal of Stroke. 2021;16(2):163–171. In 89 patients scanned within 24 hours, brain bleeds measured 57.5 ± 3.3 HU on average. Supports the fresh-bleed value in sections 2 and 3.
- Tejani AS, Klontzas ME, Gatti AA, et al. Checklist for Artificial Intelligence in Medical Imaging (CLAIM): 2024 Update. Radiology: Artificial Intelligence. 2024;6(4):e240300. Covers acquisition, preprocessing, intended use, internal and external testing, and prospective evidence. Supports the validation expectations in section 8.
- Efthimiou O, Seo M, Chalkou K, Debray T, Egger M, Salanti G. Developing clinical prediction models: a step-by-step guide. BMJ. 2024;386:e078276. Discusses discrimination, calibration, external validation, and transportability. Supports the section 6 point that a score is only a probability if the model was calibrated and validated to be one.
- Arun N, Gaw N, Singh P, et al. Assessing the Trustworthiness of Saliency Maps for Localizing Abnormalities in Medical Imaging. Radiology: Artificial Intelligence. 2021;3(6):e200267. Evaluates saliency maps on localization utility, sensitivity to model weights, repeatability, and reproducibility, and recommends detection or segmentation models where localization is the goal. Supports the heatmap caution in sections 5 and 8.
- US Food and Drug Administration. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations. Draft guidance, January 2025. Addresses intended-use populations, multisite testing, human-device interaction, bias, workflow, and lifecycle performance monitoring. Supports the acquisition-variation list in section 7 and the oversight point in section 8.
- US Food and Drug Administration. De Novo decision summary DEN170073, Viz.ai ContaCT. 2018. Describes a notification-only tool that analyzes CT angiography for a suspected large vessel occlusion, alerts a specialist based on a fixed threshold, and runs in parallel with standard care, in which radiologists read all images. Its algorithm summary states that skull stripping and registration steps are performed, and that the software excludes studies that are incomplete, of unacceptable technical quality, or show metallic artifacts such as aneurysm clips. Supports the triage framing in sections 1, 6, and 8, the skull-stripping note in section 4, and the skipped-scan caution in section 8.
- Radiopaedia. Ischaemic stroke. Notes that noncontrast CT has limited sensitivity for ischemic stroke in the first hours and is often normal or only subtly abnormal early on. Supports the note in section 1 and the third caution in section 8.
- V H, Athray B, Murthy LT, et al. Deep learning for intracranial hemorrhage detection and classification in brain CT scans: a systematic review and hybrid model approach. Frontiers in Digital Health. 2026;8:1703634. Describes Hounsfield unit windowing and multi-window input in published bleed-detection models, and names skull stripping as one optional preprocessing step. Supports the skull-stripping note in section 4.
- Budzyń K, et al. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterology & Hepatology. 2025;10(10):896–903. Nineteen experienced endoscopists at four Polish centers; adenoma detection in colonoscopies done without AI fell from 28.4% in the three months before AI came into routine use to 22.4% in the three months after. Observational, so it shows an association. Supports the deskilling note in section 8.
- Okumura T, et al. Long-term impact of computer-aided adenoma detection: a prospective observational study. Endoscopy. 2026;58(2):121–129. Single-center prospective study of 4,712 propensity-matched patients, 2,356 of them examined without AI, in which detection without AI was maintained. Supports the note in section 8 that the evidence is mixed.
- Larsen DP, Butler AC, Roediger HL. Repeated testing improves long-term retention relative to repeated study: a randomised controlled trial. Medical Education. 2009;43:1174–1181. Forty residents; repeated testing with feedback produced better knowledge retention about six months later than repeated study of the same material (39% vs 26%). Applying it to AI-related skill loss is an inference. Supports the first practice in the deskilling note.
- Buçinca Z, Malaya MB, Gajos KZ. To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proceedings of the ACM on Human-Computer Interaction. 2021;5(CSCW1):188. Experiment with 199 online participants on a nonclinical meal-planning task. Three cognitive forcing designs (answer first, wait for the advice, advice on demand) reduced overreliance compared with simple explainable AI, and people gave the least favorable ratings to the designs that reduced overreliance most. Supports the first practice in the deskilling note.
- Centers for Disease Control and Prevention. Stroke Facts. States that about 87% of all strokes are ischemic, meaning blood flow to the brain is blocked. Supports the third caution in section 8.
Where to go next
The same threshold trade-offs, in a risk score
Predictive AI, in depth follows one synthetic deterioration model through a moving threshold, a 1,000-encounter matrix, and a monitoring decision. This page took the model's weights as given; how a model learns them is the subject of a planned Training material.