🩺 Medical Scan Classifier — How a CNN Reads an Image

The ROC (receiver operating characteristic) curve plots true positive rate (sensitivity) against false positive rate (1 − specificity) as the decision threshold sweeps from 1 down to 0. Each point on the curve corresponds to one possible threshold. A random guesser traces the diagonal line; a perfect classifier hugs the top-left corner. The area under the curve (AUC) summarises performance across all thresholds at once — an AUC of 0.5 means no better than chance, 1.0 means perfect separation. Because this simulation adds random per-case noise and randomises the blob's amplitude and position, the curve you see is realistically imperfect rather than a trivial step function.

No — and that is deliberate. The kernels used here are fixed, hand-chosen weights, not the result of gradient descent on labelled training data. A real diagnostic CNN would learn its convolution weights (often across dozens of layers and hundreds of channels) by backpropagating a loss computed against thousands of expert-labelled scans. What this simulation does reproduce faithfully is the forward pass: convolution → feature map → pooling → sigmoid → probability, which is the same computational pipeline a trained network executes at inference time, just with weights chosen by hand instead of learned.

Real imaging sensors always have background noise, and real anatomy has natural texture even when healthy. Modelling "Normal" as pure noise around a low baseline intensity (rather than a blank grid) forces the classifier to distinguish a genuine elevated-intensity structure from ordinary pixel-to-pixel variation — which is the actual challenge medical image classifiers face, and the reason a naive brightness threshold on raw pixels performs far worse than a convolution-based feature detector.

The confusion matrix and sensitivity/specificity/accuracy figures are recomputed instantly whenever you move the threshold slider, using the diagnosis probability already stored for every case run so far. Only the underlying probabilities require a batch run (or single case) to generate; classifying them as positive or negative against a given cutoff is just a comparison, so it can be recalculated live without regenerating any scans.

Each case redraws its blob's amplitude, position and width from a random distribution around the preset's typical values, and adds fresh pixel noise. This mirrors real biological and imaging variability — no two real scans of the same pathology look identical. It's also what makes the ROC curve meaningful: if every "Early" case scored identically, sensitivity and specificity would jump in discrete steps rather than tracing a smooth curve.