A convolutional neural network (CNN) trained for binary tumour-vs-no-tumour classification never "sees" a brain the way a radiologist does. It slides small filters across the image, builds up feature maps that respond to edges and blobs, repeatedly pools those maps down to a coarser grid, and finally collapses everything into a single number: the estimated probability that a tumour is present.
In an NHS triage setting, a missed tumour (a false negative) is far more costly than a false alarm that a radiologist quickly rules out. Because most scans are healthy, a model can score very high "accuracy" while still missing a worrying share of real tumours — which is why teams tune the threshold to protect recall (catching true cases) even at the cost of some precision.
A synthetic brain MRI slice flows through two computed convolution-and-pool layers into a live probability gauge, so you can see how a binary tumour classifier turns pixels into a triage decision.
Each feature-map grid is computed from the actual synthetic scan using edge/blob-style filters and max-pooling, so brighter cells genuinely mark where the network finds tumour-like structure, not a scripted animation.
Pick a tumour or healthy scan, adjust tumour size and scan noise, then set the decision threshold. Run batches of scans and watch recall, precision and accuracy update — notice how a stricter threshold protects precision but can cost recall.
Because most scans in a screening population are healthy, a model can post high "accuracy" while quietly missing many true tumours — which is why NHS-style triage tools are tuned around recall (sensitivity), not accuracy alone.