Top diagram is the end-to-end flow. Below it, three image-based animated panels show what actually happens to the pixels inside Convolution+Activation, the GRU across CT slices, and Dense+Softmax. All feature maps are computed by real convolution on the de-identified nodule ROI.
Four learned-style kernels run simultaneously and continuously. In each row the 3×3 window sweeps the same input ROI, computing the dot product Σ(w·x) at every position; the raw feature map fills in, then ReLU keeps only positive responses. Notice how each filter isolates a different structure — vertical edges, horizontal edges, blob/margin, and fine texture.
A nodule spans several adjacent axial slices. Each slice is convolved into its own feature vector, then fed into the GRU in order. The hidden state h updates at every step (update & reset gates decide what to keep vs. forget), carrying 3D context forward. The final h is a whole-lesion summary.
The final hidden state is multiplied by the dense layer weights to produce two raw logits (benign score, malignant score). Softmax exponentiates and normalizes them into probabilities that sum to 1. Drag to see how relative logits become a calibrated probability.
In your pipeline this becomes a structured Finding (class + probability + criterion_version), emitted across the PHI boundary — never raw DICOM.