Patient-specific placental vessel segmentation with limited data

• DOI: 10.1007/s11701-024-01981-z
• Publication Date: 2024-06-04T13:10:44Z
• Authors: Gary Sarwin; Jonas Lussi; Simone Gervasoni; Ueli Moehrlen; Nicole Ochsenbein; Bradley J. Nelson
• Keywords: Generative adversarial networks; Research; Placenta; Twin-to-twin transfusion syndrome; 610 Medicine & health; Fetofetal Transfusion; Medical image generation; 2746 Surgery; Machine Learning; 03 medical and health sciences; Deep Learning; Segmentation; 0302 clinical medicine; Robotic Surgical Procedures; Pregnancy; Image Processing, Computer-Assisted; Humans; Generative adversarial networks; Medical image generation; Segmentation; Twin-to-twin transfusion syndrome; Female; 10220 Clinic for Surgery; Neural Networks, Computer; 2718 Health Informatics

Abstract A major obstacle in applying machine learning for medical fields is the disparity between data distribution of training images and encountered clinics. This phenomenon can be explained by inconsistent acquisition techniques large variations across patient spectrum. The result poor translation trained models to clinic, which limits their implementation practice. Patient-specific networks could provide a potential solution. Although patient-specific approaches are usually infeasible because expenses associated with on-the-fly labeling, use generative adversarial enables this approach. study proposes approach based on networks. In presented pipeline, user trains segmentation network extremely limited supplemented artificial samples generated models. demonstrated endoscopic video captured during fetoscopic laser coagulation, procedure used treating twin-to-twin transfusion syndrome ablating placental blood vessels. Compared standard deep approach, pipeline was able achieve an intersection over union score 0.60 using only 20 annotated compared 100 Furthermore, without achieves 0.30, which, therefore, corresponds 100% increase performance when incorporating pipeline. GANs generate supplements real data, allows network. We show that significantly improve vessel viable solution bring automated clinic.