Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

T2ID-CAS: Diffusion Model and Class Aware Sampling to Mitigate Class Imbalance in Neck Ultrasound Anatomical Landmark Detection

Manikanta Varaganti, Amulya Vankayalapati, Nour Awad, Gregory R. Dion, Laura J. Brattain

TL;DR

This work targets class imbalance in neck ultrasound anatomical landmark detection by introducing T2ID-CAS, a hybrid framework that fuses text-to-image diffusion (SDXL) with LoRA fine-tuning and class-aware sampling to synthesize minority-class data. The synthetic data, combined with CAS during training of a YOLOv9s detector, yields substantial improvements in mean Average Precision—particularly for underrepresented structures like tracheal rings and vocal folds—over a strong baseline. Across extensive experiments, SDXL-generated samples show improved realism, diversity, and semantic alignment, while T2ID-CAS delivers the highest overall performance (mAP across IoU thresholds), highlighting the approach's scalability and practicality for ultrasound-guided airway management. The study also discusses limitations, such as computational cost of diffusion model fine-tuning and sensitivity to prompts, and outlines directions for larger datasets and prompt robustness analyses.

Abstract

Neck ultrasound (US) plays a vital role in airway management by providing non-invasive, real-time imaging that enables rapid and precise interventions. Deep learning-based anatomical landmark detection in neck US can further facilitate procedural efficiency. However, class imbalance within datasets, where key structures like tracheal rings and vocal folds are underrepresented, presents significant challenges for object detection models. To address this, we propose T2ID-CAS, a hybrid approach that combines a text-to-image latent diffusion model with class-aware sampling to generate high-quality synthetic samples for underrepresented classes. This approach, rarely explored in the ultrasound domain, improves the representation of minority classes. Experimental results using YOLOv9 for anatomical landmark detection in neck US demonstrated that T2ID-CAS achieved a mean Average Precision of 88.2, significantly surpassing the baseline of 66. This highlights its potential as a computationally efficient and scalable solution for mitigating class imbalance in AI-assisted ultrasound-guided interventions.

T2ID-CAS: Diffusion Model and Class Aware Sampling to Mitigate Class Imbalance in Neck Ultrasound Anatomical Landmark Detection

TL;DR

This work targets class imbalance in neck ultrasound anatomical landmark detection by introducing T2ID-CAS, a hybrid framework that fuses text-to-image diffusion (SDXL) with LoRA fine-tuning and class-aware sampling to synthesize minority-class data. The synthetic data, combined with CAS during training of a YOLOv9s detector, yields substantial improvements in mean Average Precision—particularly for underrepresented structures like tracheal rings and vocal folds—over a strong baseline. Across extensive experiments, SDXL-generated samples show improved realism, diversity, and semantic alignment, while T2ID-CAS delivers the highest overall performance (mAP across IoU thresholds), highlighting the approach's scalability and practicality for ultrasound-guided airway management. The study also discusses limitations, such as computational cost of diffusion model fine-tuning and sensitivity to prompts, and outlines directions for larger datasets and prompt robustness analyses.

Abstract

Neck ultrasound (US) plays a vital role in airway management by providing non-invasive, real-time imaging that enables rapid and precise interventions. Deep learning-based anatomical landmark detection in neck US can further facilitate procedural efficiency. However, class imbalance within datasets, where key structures like tracheal rings and vocal folds are underrepresented, presents significant challenges for object detection models. To address this, we propose T2ID-CAS, a hybrid approach that combines a text-to-image latent diffusion model with class-aware sampling to generate high-quality synthetic samples for underrepresented classes. This approach, rarely explored in the ultrasound domain, improves the representation of minority classes. Experimental results using YOLOv9 for anatomical landmark detection in neck US demonstrated that T2ID-CAS achieved a mean Average Precision of 88.2, significantly surpassing the baseline of 66. This highlights its potential as a computationally efficient and scalable solution for mitigating class imbalance in AI-assisted ultrasound-guided interventions.
Paper Structure (16 sections, 4 figures, 2 tables)

This paper contains 16 sections, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Proposed Methods
  3. Text-to-Image Latent Diffusion Model
  4. Class-Aware Sampling
  5. Neck anatomical landmark detection
  6. Materials and Experiments
  7. Data Description
  8. Training Configuration
  9. Image Generation with Stable Diffusion XL
  10. Comparative Experiments
  11. Results and Discussions
  12. Stable Diffusion XL results
  13. YOLOv9s results
  14. Discussions
  15. Conclusion
  16. ...and 1 more sections

Figures (4)

  • Figure 1: Overview of the proposed T2ID-CAS framework
  • Figure 2: Long-tailed distribution of instance and image counts per class in the neck US dataset. Colored bars indicate the number of annotated instances (blue) and the number of images (orange) per class.
  • Figure 3: Comparison between original images and synthetic images by SDXL. (Top) shows the results of tracheal ring using prompt "Ultrasound image of human tracheal ring". (Bottom) shows the results of vocal fold using prompt "Ultrasound image of human vocal fold"
  • Figure 4: Precision-Recall curves for Tracheal Ring (Top) and Vocal Fold (Bottom) across different configurations.