A scan image is fed into the input layer and propagates forward
through hidden layers, each node combining weighted signals from the
previous layer and firing through an activation function, until the
output layer produces a single diagnostic confidence score.
a_l = f(W_l · a_(l-1) + b_l) (forward propagation, layer by layer)
confidence = a_output
diagnosis = "positive" if confidence ≥ threshold
- Network depth — number of hidden layers; deeper networks can learn more complex patterns but also propagate noise further.
- Confidence threshold — the cutoff above which the model calls a case positive; raising it reduces false positives but may miss real cases.
- Input noise level — scan artefacts or poor image quality that corrupt the input signal before it ever reaches layer one.
Real clinical-grade models are validated far more rigorously than
this toy demo, but the forward-pass mechanism — weighted sums,
activation, layer-by-layer propagation — is exactly the same idea.