StrokeDetector: A lightweight multi-output deep learning pipeline for ischaemic stroke analysis on diffusion-weighted MRI

StrokeDetector is an open-source Python package for automated acute ischaemic stroke analysis from diffusion-weighted MRI (DWI) and apparent diffusion coefficient (ADC) maps. Unlike existing tools that output only voxel-level binary segmentation masks, StrokeDetector returns five outputs from a single inference pass: a stroke presence flag, a 3D bounding box with centre coordinates, a lesion morphology subtype label, a voxel-level segmentation mask in native image space, and infarct volume in millilitres. The pipeline is a 0.504-million-parameter 3D ResNet with a Feature Pyramid Network, trained on ISLES-2022 using a three-stage sequential strategy that separates gradient-conflicting detection and presence-discrimination objectives. On the 36-case held-out ISLES-2022 test set the pipeline achieves mAP@0.2–0.5 = 0.798, mean Dice on positive cases = 0.634, sensitivity = 0.905, F1 = 0.731, and runs in under ten seconds on a single NVIDIA T400 4 GB GPU. All outputs are returned in native image space and load directly in standard viewers without co-registration.