Durgical
EN
Open the app

AI in medical imaging: convolutional neural networks

  • Concept

Most medical AI that is already in use reads images. Knowing how it sees a scan tells you when to trust its flag and when to look again yourself.

Updated

What it is

A convolutional neural network (CNN) is a deep learning model built for images. Small learned filters slide across the pixels and build features layer by layer, from edges to shapes to whole patterns, and the last layers turn them into an output such as a probability of disease, a box around a finding or an outline of an organ. It learns from thousands of labelled images rather than from rules written by people. In practice it is a diagnostic test that sits inside software, so it is judged, authorised and monitored as one.

Key facts

  • A filter is a small grid of weights, for example 3 × 3 pixels; sliding it across the image gives a feature map that lights up where its pattern appears.
  • Pooling shrinks the feature maps, so each deeper layer sees a wider area: edges become textures, textures become a nodule or a fracture line.
  • Training adjusts the weights to reduce the error between the output and the label, repeated over many images; nobody tells the model which features to use.
  • Heatmaps show which pixels drove the output; they help spot a wrong cue but do not prove the model reasoned correctly.
  • Locked algorithms give the same output for the same image after authorisation; adaptive ones change with new data, so changes need a pre-agreed change control plan.
  • Software intended for a medical purpose that runs without being part of a hardware device is called software as a medical device (SaMD).

Worked example

  1. A hospital tests a head CT bleed triage model on 1 000 consecutive studies: 80 (8 %) show intracranial haemorrhage, 920 do not.
  2. Operating point A (low threshold): sensitivity 95 %, specificity 85 %. It flags 76 of 80 bleeds and 15 % of 920 = 138 scans without a bleed.
  3. Point A: 76 + 138 = 214 flags; PPV = 76 / 214 ≈ 36 %; 4 bleeds are not flagged and wait in the normal queue.
  4. Operating point B (higher threshold): sensitivity 85 %, specificity 95 %. It flags 68 bleeds and 5 % of 920 = 46 scans without a bleed.
  5. Point B: 68 + 46 = 114 flags; PPV = 68 / 114 ≈ 60 %; 12 bleeds are not flagged.
  6. Moving from A to B removes 92 false flags per 1 000 studies but leaves 8 more bleeds without priority.
  7. A radiologist still reads every study, so a missed flag means a delay, not a missed diagnosis: for triage, point A is usually the better choice.
  8. Reading: the operating point is a clinical decision on the ROC curve, weighing the cost of delay against false alarms at local prevalence.

Imaging AI tasks: what each one answers

TaskQuestion it answersOutputExampleHow it is judged
ClassificationIs the finding present?A probability or class for the whole imageReferable diabetic retinopathy on a retinal photographSensitivity, specificity, ROC curve
Detection or localisationWhere is it?A box or point on each findingLung nodules on chest CT, fractures on radiographsFindings found and false marks per image
SegmentationWhat is its exact outline?A pixel-by-pixel maskOutline of a tumour, an infarct or the heart chambersOverlap between the model's mask and the expert's mask
TriageWhich study should be read first?A flag that moves the study up the worklistLarge-vessel occlusion or haemorrhage on head CTTime to read and to treatment; false flags per day
QuantificationHow big, how much, how fast is it changing?A number such as a volume or a densityNodule volume growth, breast density, ejection fractionAgreement with expert measurement and repeatability

Markers

  • CNN
  • SaMD
  • AUC
  • PCCP

In practice

  • Diabetic retinopathy screening: a model grades retinal photographs in primary care and refers positive patients to an eye specialist.
  • Mammography: a model marks suspicious areas or acts as a second reader, aiming to keep cancer detection with fewer recalls.
  • Chest radiographs and CT: models flag nodules, pneumothorax or consolidation and measure nodule growth between scans.
  • Stroke and haemorrhage: on head CT or CT angiography a triage model alerts the team to a possible large-vessel occlusion or bleed.
  • Fracture detection: in the emergency department a model marks subtle fractures on radiographs for the clinician to check.
  • Before relying on a tool, ask: which task, which scanners and population it was tested on, at which operating point, and who reviews its errors.

Common pitfalls

  • Reading a model's AUC from the developer's data as if it held on your scanners and protocols.
  • Taking a reader study on an enriched case mix as proof of benefit in daily practice.
  • Not counting the false-flag load per day before switching on a triage tool.
  • Skipping monitoring after installation, so a software update or a new scanner silently lowers performance.

Common traps

  • The CNN looks for disease like a radiologist: it learns any pattern that separates the labels, including text markers, rulers, a portable-film label or the hospital's scanner, so it can be right for the wrong reason.
  • Beating radiologists in a reader study proves benefit: reader studies use enriched case sets and artificial conditions; real prevalence, prospective use and the human with the tool decide the effect.
  • No flag from a triage tool means the study is normal: triage only reorders the worklist; an unflagged study still needs a full read.
  • The training labels are the truth: labels from a single reader carry that reader's misses; a better reference standard uses consensus, follow-up or pathology.
  • An authorised algorithm keeps learning safely on its own: a locked model does not change, and changes to an adaptive one must be reviewed or pre-agreed in a change control plan; either still needs local monitoring.

Test yourself

Think of the answer, then tap the question to see it.

A fracture model performs well at its home hospital but poorly on images from a new vendor's machines. Why, and what should be done?

Dataset shift: different scanners and protocols change the pixels the model relies on. Test it on local images before use and monitor it after deployment.

A chest radiograph model detects pneumothorax mainly in films that already show a chest drain. What has it learned?

A spurious cue: the drain marks treated cases. It may miss untreated pneumothorax, the cases that matter; test it on images without drains.

For a haemorrhage triage tool where every CT is still read, why choose a high-sensitivity operating point?

A missed flag only delays a read, while a caught bleed is read sooner; the price is more false flags, which should be counted per 1 000 studies.

Why does a reader study on an enriched set of 50 % cancers overstate the PPV a radiologist will see in screening?

PPV rises with prevalence; screening prevalence is far lower, so the same sensitivity and specificity give many more false positives per true one.

A radiologist changes a correct read to agree with the AI's wrong mark. What is this, and what reduces it?

Automation bias. Reading first and then viewing the AI output, training on the tool's typical errors and auditing disagreements reduce it.

Add to my cards (5)

Elsewhere in DurgiPedia

Connections

← All of epidemiology

Educational content on epidemiological and statistical methods; the examples use made-up numbers to teach the calculation. Not a clinical or legal guide: current Colombian rules are in the official sources.

Sources: Artificial Intelligence-Enabled Medical Devices (U.S. Food and Drug Administration, 2026); List of Artificial Intelligence-Enabled Medical Devices (U.S. Food and Drug Administration, 2026); Software as a Medical Device (SaMD) (U.S. Food and Drug Administration, 2018); Ethics and governance of artificial intelligence for health (World Health Organization, 2021).

Spotted a mistake?We are always improving for the community. Tell us what is wrong and our team will review it.Report a mistake

Representative model, for educational purposes only.