Advanced semantic segmentation using deep learning

Abstract

[en] Intravascular Ultrasound (IVUS) is an intra-operative imaging modality that facilitates observing the vessel wall structure of the human coronary arteries. Segmentation of the arterial wall from IVUS images is crucial for the analysis of the wall’s characteristics and for 3D reconstructed models of the artery. The aim of this project is to learn, study, develop and evaluate a model capable of performing semantic segmentation on medical images using a small IVUS dataset. The dataset used contains 435 pullbacks from 10 different patients acquired by 20 MHz probes. Our proposed model uses the InceptionResNet architecture pretrained with the ImageNet dataset as the encoder. The model is then connected like a U-Net with a decoder which is trained using the images. We have explored different approaches in data augmentation and the best results were given by rotations up to 90 degrees, vertical & horizontal flips and zooming. Among the results, we obtained an average of Dice similarity coefficient, or Dice Index, of 0.75 for the media and 0.92 for the lumen, with pretty low variance. However, the Hausdorff distance (HD) from results show that some misclassified sections appear in average 0.51 and 0.67 mm away from the ground truth class, for the lumen and media respectively. Finally, the Jaccard Measure for the lumen is quite high: 0.85; whereas for the media is 0.60, which is pretty low. The results found in this project present a good start on how to implement a solution for the problem explained.