BS: Take the Hint - Interactive Multitracer PET/CT Lesion Segmentation with a Scribble-Conditioned ResEnc U-Net
A scribble-conditioned residual encoder U-Net for interactive multitracer PET/CT lesion segmentation was submitted to the autoPET/CT V challenge. The model uses four input channels: CT, PET, and sparse scribble maps for foreground and background. It is initialized from autoPET-III winning weights, with scribble channels zero-initialized to preserve pretrained representation. PET intensities are normalized using a per-scan aorta blood-pool reference from CT segmentation. Five fold models are ensembled by averaging at inference.
The autoPET/CT V challenge addresses automated lesion segmentation in whole-body PET/CT by making it interactive, providing user scribbles marking foreground and background. The submission is a scribble-conditioned residual encoder U-Net with four input channels: CT, PET, and sparse scribble maps for foreground and background. The network is initialized from autoPET-III winning weights and extended from two to four input channels, with scribble channels zero-initialized to preserve the pretrained representation exactly. Each model is fine-tuned per fold from the corresponding autoPET-III fold checkpoint, ensuring no validation case is seen during pretraining. PET intensities are normalized against a per-scan aorta blood-pool reference derived from a CT segmentation, removing tracer- and centre-specific scaling without requiring lesion labels. At inference, five fold models are ensembled by averaging.
The zero-initialization of new scribble input channels allows the pretrained autoPET-III weights to be preserved exactly at initialization, enabling fine-tuning without catastrophic forgetting. Per-scan aorta blood-pool normalization provides a robust intensity reference that is independent of lesion labels, improving generalization across tracers and centers. Ensembling five fold models by averaging likely reduces variance and improves segmentation robustness.
Interactive segmentation with scribbles reduces the annotation burden for clinicians and can accelerate the adoption of AI-assisted lesion detection in PET/CT workflows. The approach leverages existing winning models (autoPET-III) as a foundation, demonstrating transfer learning across challenge iterations. Normalization based on anatomical references (aorta) may facilitate deployment across diverse clinical sites without retraining.
The method could be integrated into medical imaging software to provide interactive, semi-automated lesion segmentation, reducing manual contouring time and improving consistency. It may lower barriers to AI adoption in nuclear medicine by requiring only simple scribbles rather than full annotations. The use of pretrained models and efficient fine-tuning suggests lower development costs for vendors.
Next observable signals include the official autoPET/CT V challenge results and any follow-up publications or open-source code releases. Potential extensions could incorporate uncertainty estimation from the ensemble or adapt the scribble conditioning to other imaging modalities. Clinical validation studies may assess the impact on radiologist efficiency and diagnostic accuracy.