Placental Vessel Segmentation Using Pix2pix Compared to U-Net.

Area of Science:

• Medical Imaging
• Computer-Assisted Surgery
• Deep Learning

Background:

• Fetoscopic laser surgery relies on accurate placental vessel segmentation.
• Current segmentation methods exhibit significant intra- and inter-procedure variability.
• Addressing this variability is crucial for improving surgical outcomes.

Purpose of the Study:

• To compare the performance of conditional generative adversarial networks (cGANs) against the U-Net model for placental vessel segmentation in fetoscopic images.
• To evaluate the efficacy of the pix2pix cGAN model for this specific application.
• To assess the potential of deep learning to reduce segmentation variability in computer-assisted fetoscopic surgery.

Main Methods:

• Trained and evaluated two deep learning models: U-Net and pix2pix (a cGAN).
• Utilized both a publicly available dataset and an internal validation set for comprehensive testing.
• Quantitatively compared segmentation performance using Dice and Intersection over Union (IoU) scores.

Main Results:

• The pix2pix model demonstrated superior performance over U-Net on the public dataset (Dice: 0.80 vs 0.75, IoU: 0.70 vs 0.66).
• Internal validation confirmed pix2pix's advantage (Dice: 0.68 vs 0.53, IoU: 0.59 vs 0.49).
• Both metrics showed statistically significant improvements with the cGAN approach (p < 0.01).

Conclusions:

• Conditional generative adversarial networks, specifically pix2pix, offer improved placental vessel segmentation in fetoscopic images compared to U-Net.
• The cGAN-based approach shows promise in enhancing the precision of computer-assisted fetoscopic surgery.
• Further investigation is needed to address the generalizability of these models across diverse surgical scenarios.