T2ID-CAS: Diffusion Model and Class Aware Sampling to Mitigate Class Imbalance in Neck Ultrasound Anatomical Landmark Detection (2504.21231v1)

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.

Mitigating Class Imbalance in Ultrasound Imaging with T2ID-CAS

The paper "T2ID-CAS: Diffusion Model and Class Aware Sampling to Mitigate Class Imbalance in Neck Ultrasound Anatomical Landmark Detection" addresses the prevalent issue of class imbalance in medical imaging datasets, specifically focusing on the detection of anatomical landmarks in neck ultrasound (US) imaging. This imbalance often poses a significant barrier to the effective training of deep learning models due to certain anatomical structures, like tracheal rings and vocal folds, being severely underrepresented.

The paper introduces an innovative approach, T2ID-CAS, which combines a text-to-image latent diffusion model and class-aware sampling (CAS) to improve the representation of minority classes in the dataset. The core of the methodology lies in generating high-quality synthetic images for these underrepresented classes using a fine-tuned Stable Diffusion XL model (SDXL) that is conditioned on textual prompts. This inclusion of synthetic data alongside CAS effectively balances the training dataset, leading to superior model performance in anatomical landmark detection tasks.

Crucially, the paper outlines the technical contributions of this approach:

1. Fine-tuning SDXL: The paper describes the process of fine-tuning a pre-trained SDXL model specifically for neck ultrasound imaging using Low-Rank Adaptation (LoRA) to enhance computational efficiency. The model generates realistic and diverse synthetic images of minority classes by responding to class-specific textual prompts.
2. Utilization of Class-Aware Sampling: The integration of CAS provides an adaptive sampling mechanism that ensures an even representation of all classes within training mini-batches, enhancing the model's ability to learn features relevant to underrepresented classes.
3. Performance Measurement: The paper uses a YOLOv9 based object detection model to demonstrate the effectiveness of the proposed approach. The results indicate a significant improvement in mean Average Precision (mAP50-95), achieving 88.2% as compared to a baseline of 66%, indicating substantial gains in detection accuracy across all anatomical classes.

The numerical results, particularly the mAP gains seen for tracheal rings and vocal folds, underscore the effectiveness of integrating synthetic data for previously underrepresented classes. The capacity to generate these synthetic US images, evaluated with metrics such as Fréchet Inception Distance (FID) and Inception Score (IS), reveals the high fidelity and diversity of the generated data, fostering improved generalization in the detection models.

From a theoretical perspective, this research underscores the potential of leveraging advanced generative models like SDXL in medical imaging, especially in domains where class imbalance is severe. The application of diffusion models in ultrasound imaging, a relatively unexplored area, opens avenues for addressing data scarcity and imbalance, which are common hurdles in clinical settings.

On a practical level, the improved detection accuracy facilitated by T2ID-CAS can enhance the reliability of ultrasound-guided procedures in clinical practice. This is particularly crucial for rapid and accurate airway management, potentially reducing risks associated with misplacements and improving patient outcomes, especially in critical care scenarios.

Looking forward, future research could explore expanding this framework to other medical imaging domains characterized by class imbalance. Additionally, further refinement of computational strategies for generative models could yield even more efficient and scalable solutions in this space. The intersection of generative AI and medical imaging thus represents a fertile ground for advancing both clinical and technological frontiers.