ASTRA-Net: Anatomy-Specific Transfer and Representation Alignment for Drug-Induced Sleep Endoscopy Segmentation

2026-07-23 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition, Medical Devices & Health Technology · Depth: Expert, quick

Summary

ASTRA-Net is a novel deep learning framework designed for anatomy-specific segmentation in drug-induced sleep endoscopy (DISE) when real annotations are limited. It employs a two-stage process. Stage 1 aligns intermediate ConvNeXt-Base representations from 14,250 unlabeled virtual endoscopy frames, derived from CT scans, with real DISE frames for feature alignment. Stage 2 then fine-tunes four independent UNet++ decoders using 401 real annotated frames, incorporating structured zero-mask supervision. On a 100-frame hold-out evaluation set, a five-model MMD-only segmentation ensemble achieved a mean Dice score of 0.8927, with a 95% image-level bootstrap interval of 0.8631 to 0.9160, and a mean intersection over union of 0.8239. A classification-enabled variant also reached a 0.92 top-1 accuracy for restricted four-plane identification.

Key takeaway

For machine learning engineers developing medical image segmentation models with scarce annotated data, ASTRA-Net offers a robust strategy. You should consider integrating virtual data for representation alignment and employing a two-stage fine-tuning process with structured supervision. This approach can significantly improve segmentation accuracy, as demonstrated by the 0.8927 Dice score, even when real annotations are severely limited.

Key insights

ASTRA-Net enables accurate DISE segmentation with limited real data by aligning virtual and real image representations.

Principles

• Feature alignment improves segmentation with scarce data.
• Virtual data can enhance real-world model performance.
• Ensembling models boosts segmentation reliability.

Method

ASTRA-Net uses a two-stage approach: first, align ConvNeXt-Base features from virtual and real DISE frames, then fine-tune UNet++ decoders on limited real annotations with zero-mask supervision.

In practice

• Utilize CT-derived virtual endoscopy frames.
• Implement MMD for domain alignment.
• Employ UNet++ decoders for fine-tuning.

Topics

• Drug-Induced Sleep Endoscopy
• Medical Image Segmentation
• Domain Adaptation
• Representation Alignment
• UNet++
• ConvNeXt-Base
• Limited Annotation Learning

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Multi-layer feature aggregation network with residual module and attention mechanism for jaw cyst image segmentation — MFA-Net enhances medical image segmentation by integrating residual modules, channel-spatial attention, and multi-layer feature aggregation.
• UnDA: Unpaired Domain Alignment for Cross-Modal Knowledge Transfer in Medical Imaging — UnDA enables robust cross-modal knowledge transfer in medical imaging using unpaired data by mitigating noise and aligning features.
• Domain Shift in Echocardiography: Interpretable Quantification and Prediction of Cross-Dataset Left Ventricular Segmentation — Echocardiographic domain shift is measurable and its impact on LV segmentation can be reduced by geometry-aware preprocessing and predicted using representation-specific metrics.
• ConvNeXt-FD: A Fractal-Based Deep Model for Robust Biomedical Image Segmentation — ConvNeXt-FD enhances biomedical image segmentation by combining ConvNeXt with fractal-inspired boundary regularization for superior shape fidelity.
• Multi-Stage Bi-Atrial Segmentation Framework from 3D Late Gadolinium-Enhanced MRI using V-Net Family Models — A multi-stage V-Net framework with MCLAHE preprocessing effectively segments bi-atrial regions from 3D LGE-MRI.
• Embracing Intra-Class Heterogeneity for Semi-Supervised Medical Image Segmentation: From Diversity to Precision — Explicitly modeling intra-class heterogeneity with multiple intensity-aligned prototypes improves semi-supervised medical image segmentation precision, especially with limited data.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.