🩻 Beyond a Single CNN

About this simulation

A single convolutional neural network trained to flag pneumonia on a chest X-ray gives one number and no sense of how much to trust it. This scene shows a more clinically honest pipeline: four independently pretrained CNNs each score the same film, their outputs are combined into a weighted ensemble, Monte Carlo Dropout is used to resample the network many times to see how much the prediction wobbles, and a decision threshold — tuned with Youden's J statistic on a validation ROC curve — turns the noisy probability into a clear/flag call.

How it works

• Model bank — four stylised CNN "readers" (DenseNet121, ResNet50, EfficientNet-B0, VGG16) each render a probability bar for the current film; beams carry each model's vote to the central ensemble node, with beam thickness showing its weight.
• Ensemble node — the glowing sphere is the weighted average probability, dropped as a marker onto the 0–1 probability track.
• Uncertainty cloud — the scattered spheres are repeated Monte Carlo Dropout forward passes through the same ensemble; their spread along the track is the model's own estimate of its uncertainty (σ).
• Threshold plane — the green/red split on the track is the decision threshold. The right-hand panel plots a synthetic validation ROC curve with a marker at the current threshold and Youden's J = sensitivity + specificity − 1 annotated.

Controls

• Patient case — swaps the underlying film: a clear normal film, an ambiguous mild-opacity case that sits near the boundary, or an obvious severe consolidation.
• Model mode — compare a single CNN against equal-weighted and validation-weighted 4-model ensembles; ensembling narrows the uncertainty cloud.
• MC Dropout passes / Dropout rate — control how many stochastic forward passes are sampled and how aggressively dropout perturbs them, changing the size of the uncertainty cloud.
• Decision threshold — slide the cutoff along the ROC curve; "🎯 Snap to optimal J" jumps to the threshold that maximises Youden's J for this synthetic classifier.

Did you know?

Youden's J statistic (J = sensitivity + specificity − 1) was proposed by W. J. Youden in 1950 as a single-number summary of a diagnostic test's discriminative power, and the threshold that maximises it is a common, principled starting point for turning a continuous screening score into a binary clinical decision.