Image-guided intervention of coronary lesions

Fig.1 Longitudinal view of a coronary artery in IVUS. Blue zones show the detected lumen.

Petia Radeva, Simone Balocco, Francesco Ciompi, Juan Rigla
“Computer Vision and Machine Learning at the University of Barcelona” Consolidated Research Group
University of Barcelona

IntraVascular UltraSound images (IVUS) represent a unique medical imaging modality that allows for the visualization and characterization of coronary vessels and lesions in high-resolution (Fig.1). By emitting and receiving ultrasound waves from inside the coronary arteries, IVUS has the unique possibility to visualize different structures of plaque: fibrous, mild, calcium, and mix of them. Hence, IVUS is a very important tool for image-guided intervention in coronary events. Another important feature of IVUS is the fact that it is real-time intra-operative medical imaging modality. Being at the same time minimally-invasive makes the technique of high interest for the safe treatment of patients.

However, the interpretation of IVUS is not trivial. Artifacts like speckle noise, coronary movement, attenuation of the ultrasound signal, 2D image representing the intersection of the vessel instead of its 3D shape, make the interpretation of the images a very ambiguous, tedious, subjective and time-consuming process. All this leads to a clear necessity for the development of automatic methods for analysis, allowing for more objective measures and thus more reliable analysis of the images and the image-guided intervention.</p

One of the most difficult and slow processes for physicians is the segmentation of lumen as well as media/adventitia coronary layers and plaque characterization. Due to the speckled nature of the images, the problem is difficult to solve automatically. In our group, we explored a new compact and holistic representation [Ciom12] of coronary structures in order to segment them within their context (Fig.2). Recently, we have seen that the context can be very efficiently trained using a deep learning approach. Then, a Convolutional Neural Network (CNN) is used in order to characterize the different regions of plaques for the lumen and media/adventitia detection.

Fig.2 Vessel segmentation and tissue characterization

IVUS is becoming more important as one of the few intra-operative modality and tools for image interventions since it is real-time and minimally-invasive. This technique is extremely useful in order to assess the correct deployment of stents. However, as clinical practice shows there is a significant amount of cases where the stent once posed inside the vessel is not expanded enough. The empty space between stent and lumen can provoque turbulences and plaque accumulation provoking that the “remedy is worse than the diseases. In our group, we developed a new and efficient probabilistic technique [Ciom16] to detect the boundaries and assess the position of the stent along the pullback (Fig.3).

Fig.3 Stent detection and tissue characterization
  • [Ciom16] F. Ciompi, S. Balocco, J. Rigla, X. Carrillo, J. Mauri, P. Radeva “Computer-Aided Detection of Intra-Coronary Stent in Intravascular Ultrasound Sequences” Medical Physics. Volume 43, Issue 10, October 2016.
  • [Ciom12] F. Ciompi, O. Pujol, C. Gatta, M. Alberti, S. Balocco, X. Carrillo, F. Mauri, P. Radeva, “HoliMAB: a Holistic approach for Media-Adventitia border detection in Intravascular Ultrasound” Medical Image Analysis Volume 16, Issue 6, August 2012, Pages 1085–1100 (2012)