Real-time placental vessel segmentation in fetoscopic laser surgery for Twin-to-Twin Transfusion Syndrome

Szymon Płotka, Tomasz Szczepański, Paula Szenejko, Przemysław Korzeniowski, Jesus Rodriguez Calvo, Asma Khalil, Alireza Shamshirsaz, Robert Brawura-Biskupski Samaha, Ivana Išgum, Clara I. Sánchez, Arkadiusz Sitek 

Twin-to-Twin Transfusion Syndrome (TTTS) occurs in about 15% of monochorionic pregnancies, where identical twins share one placenta.Fetoscopic laser photocoagulation (FLP) is the established therapeutic approach for this condition, substantially enhancing the chances of survival for the fetuses. The procedure aims to locate and eradicate abnormal vascular connections to normalize the blood distribution between the twins. Yet, fetoscopic operations are technically demanding, primarily due to poor visibility, and significant variability among patients and domains.

To advance the visualization capabilities during these interventions, we have introduced TTTSNet, a specialized network designed to segment placental vessels accurately in real-time. This network integrates innovative elements such as a channel attention module and a feature fusion module that operates at multiple scales, ensuring detailed and precise imaging of the placental vessels, including the smaller ones. Additionally, to overcome the typical visual disturbances encountered during FLP, such as those caused by the fiberscope and amniotic debris, we implemented advanced data augmentation strategies. These strategies effectively recreate various surgical artifacts in the training data, enhancing the model’s ability to generalize across different scenarios.

TTTSNet was rigorously trained on a dataset comprising 2060 video frames from 18 distinct fetoscopic surgeries and further validated on an external dataset from 24 in-vivo procedures, which included 2348 video frames. The results were superior to those of current leading methods, achieving a mean Intersection over Union (IoU) of 78.26% across all detected vessels, and 73.35% for smaller vessels specifically. The performance rates of 172 frames per second on an A100 GPU and 152 frames per second on a Clara AGX platform demonstrate the potential of TTTSNet to support real-time surgical applications. For broader accessibility and use, the TTTSNet code has been made publicly available at the following URL: https://github.com/SanoScience/TTTSNet.

Keywords: Deep learning, Semantic segmentation, Twin-to-Twin Transfusion Syndrome (TTTS), Fetoscopic Laser Surgery

DOI: 10.1016/j.media.2024.103330