OBBSeg: Irregular Lesion Segmentation under Oriented Bounding Box Annotations

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Health & Medical Research · Depth: Expert, long

Summary

The paper introduces OBBSeg, a novel intermediate supervision paradigm for medical image segmentation. OBBSeg leverages Oriented Bounding Boxes (OBBs) to bridge the gap between weak and full supervision, offering geometry-aware annotations with low labeling cost. OBBs provide compact geometric supervision, better aligning with elongated or anisotropic lesions compared to conventional bounding boxes. To address the inherent rectangular bias of OBBs, OBBSeg incorporates a differentiable Mask-to-OBB (M2O) loss. Additionally, it integrates prompt-driven semantic guidance through two modules: Prompt-assisted Foreground Enhancer (PAFE) and Differential-based Foreground Enhancer (DBFE), which enhance foreground representation and suppress background interference. Experiments on 13 datasets across 5 imaging modalities demonstrate that OBBSeg outperforms existing weakly supervised methods and achieves performance comparable to fully supervised approaches. OBB annotation averaged 17.8 seconds per OBB, significantly less than dense mask annotation (91.1 seconds per image). The model shows robustness to OBB angle variations (±30° causing <2% drop) and padding shifts (up to 15 pixels negligible effect).

Key takeaway

For Machine Learning Engineers developing medical image segmentation models, if you are struggling with the high cost of pixel-level annotations, consider adopting an OBB-guided approach like OBBSeg. This method significantly reduces annotation effort (17.8s per OBB vs. 91.1s per mask) while achieving performance comparable to fully supervised models. You should explore integrating the Mask-to-OBB loss and prompt-guided modules to enhance geometric accuracy and foreground discrimination, especially for anisotropic lesions.

Key insights

Oriented Bounding Boxes (OBBs) combined with prompt-guided learning offer efficient, geometry-aware supervision for medical image segmentation.

Principles

• OBBs provide stronger geometric constraints for elongated lesions than axis-aligned boxes.
• Differentiable loss can mitigate rectangular bias from coarse box annotations.
• Prompt-driven semantic guidance enhances foreground representation in weak supervision.

Method

OBBSeg uses a ViT backbone with PAFE and DBFE modules for feature enhancement. It employs a dual supervision strategy: Mask-to-OBB (M2O) loss for geometry-consistent learning and Prompt Supervision Loss for hierarchical guidance.

In practice

• Use OBBs for medical image annotation to reduce cost while retaining geometric fidelity.
• Implement M2O loss to reduce rectangular bias in box-supervised segmentation.
• Integrate prompt-guided modules (PAFE, DBFE) to improve foreground-background discrimination.

Topics

• Medical Image Segmentation
• Weakly Supervised Learning
• Oriented Bounding Boxes
• Prompt-guided Learning
• Lesion Segmentation
• Mask-to-OBB Loss

Code references

• StarLxc3/OBBSeg

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

Related on AIssential

• OBBSeg: Irregular Lesion Segmentation under Oriented Bounding Box Annotations — OBBSeg uses oriented bounding boxes and semantic guidance for efficient, accurate medical image segmentation.
• Clinically Aligned Geometry Constraints for Robust IVUS Vessel Boundary Segmentation — GeoCat improves IVUS segmentation by integrating geometry consistency loss for clinical accuracy.
• Vision-Language Models as Zero-Annotation Oracles in Histopathology — General-purpose Vision-Language Models (VLMs) offer robust, zero-annotation foreground segmentation in histopathology, generalizing across diverse stains.
• Complementary Roles of Image Classification and Vessel Segmentation in AI-Based Screening for Retinopathy of Prematurity Plus Disease in a Kenyan Preterm Cohort — Image classification and vessel segmentation offer complementary strengths for AI-based ROP Plus disease detection.
• Align then Refine: Text-Guided 3D Prostate Lesion Segmentation — Text-guided, phase-scheduled training improves 3D prostate lesion segmentation by integrating semantic alignment and confidence-aware refinement.
• A deep learning framework for efficient pathology image analysis — EAGLE's selective tile analysis significantly boosts computational efficiency and predictive accuracy in digital pathology.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.