ETD

Archivio digitale delle tesi discusse presso l'Università di Pisa

Tesi etd-05202018-190514


Tipo di tesi
Tesi di laurea magistrale
Autore
GHERARDINI, MARTA
URN
etd-05202018-190514
Titolo
Enhancing Catheter Segmentation in 2D X-RayFluoroscopy Using CNNs trained on SyntheticData
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
BIONICS ENGINEERING
Relatori
relatore Prof.ssa Menciassi, Arianna
Parole chiave
  • synthetic data
  • transfer learning
  • deep learning
  • catheter tracking
  • semantic segmentation
  • medical images
  • minimally invasive surgery
Data inizio appello
07/06/2018
Consultabilità
Non consultabile
Data di rilascio
07/06/2088
Riassunto
Minimally invasive endovascular procedures require accurate tracking and localization
of tools under fluoroscopic guidance. In this work, a novel method to fullyautomatically
and in real-time segment catheters and guidewires in 2D X-Ray images
based on Deep Convolutional Neural Networks is presented. A transfer learning
approach is followed and the training process is carried out by using synthetic images
to perform the bulk of training. A small number of annotated data is then
used to fine-tune an adapted U-net model, a particular Deep Architecture that has
shown promising results in medical image segmentation tasks. The network takes as
input a single grayscale image and outputs the catheter and guidewire segmentation
within an average time of 6 ms. Two different experiments are presented in which
the network is trained on synthetic and ex-vivo fluoroscopy images, respectively. In
the latter, ex-vivo data indicate four fluoroscopic sequences acquired on a silicon
aorta phantom during catheter insertion. After the training step, by fine-tuning the
deepest layers of the network on a small amount of annotated data, accurate segmentation
performance, with an average Dice Coefficient higher than 50%, can be
obtained on in-vivo fluoroscopic images and sequences. Since the two experiments,
i.e. the training on synthetic (experiment-1) and on ex-vivo frames (experiment-
2), give comparable results, it can be argued that it is possible to reduce the need
of annotated data in the training phase. This represents an important advantage
because acquiring pixel-level annotated images is considered a key bottleneck in
building Deep Architecture segmentation models.
File