Tubular shape aware data generation for segmentation in medical imaging.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Chest X-rays are common diagnostic tools.
• Accurate segmentation of tubular structures like catheters in X-rays is crucial for interventional radiology.
• Manual annotation of medical images is time-consuming and resource-intensive.

Purpose of the Study:

• To develop a method for synthetic data generation to overcome the scarcity of annotated medical images.
• To improve the accuracy and efficiency of segmenting tube-like objects in X-ray images.

Main Methods:

• Utilized a generative adversarial network (GAN) for synthetic image generation.
• Incorporated Frangi-based regularization for shape constraints during synthetic tube generation.
• Employed an adversarial component to ensure the realistic appearance of synthesized images.

Main Results:

• The proposed method effectively generates synthetic data, eliminating the need for paired image-mask data.
• Weakly labeled datasets combined with fine-tuning on a small sample (10-20 images) achieved accuracy comparable to fully supervised models.
• Demonstrated the approach's applicability for segmenting tubes and catheters in X-ray images.

Conclusions:

• The synthetic data generation approach significantly alleviates the annotation burden in medical imaging tasks.
• The method shows promise for segmenting tubular objects across various imaging modalities and applications.
• This technique can advance the development of automated image analysis tools in radiology.