Cascade-CNN pipeline for whole-gland / central-gland / peripheral-zone segmentation from prostate MRI (T2W + DWI + ADC), built for a SoftwareX software-description paper.
This page is a static placeholder. The interactive Gradio demo (upload MRI volumes, run the real cascade pipeline, view WG/CG/PZ overlays) is built and verified locally, but is not deployed here yet — HuggingFace's free tier no longer hosts Gradio/Docker Spaces on CPU compute; that requires a PRO subscription. The code is ready to deploy as soon as compute is available.
In the meantime, the full source, trained checkpoints, and an interactive dashboard that runs locally are available on GitHub.
Held-out test split of Prostate158 (19 cases).
| Model | WG Dice | CG Dice | PZ Dice | Zone-mean |
|---|---|---|---|---|
| Cascade-CNN, direct 3-class | 0.8941 | 0.8642 | 0.7308 | 0.7975 |
| Cascade-CNN, PZ-derived — current best | 0.8926 | 0.8676 | 0.7483 | 0.8079 |
| Confidence-weighted fusion (no extra training) | 0.8910 | 0.8619 | 0.7513 | 0.8066 |
| Human inter-reader agreement (reference ceiling) | — | 0.8726 | 0.7473 | — |