Identification and localization of septal flash by the means of time distributed networks and class activation maps

Capallera Guirado, Isaac

Identification and localization of septal flash by the means of time distributed networks and class activation maps

Citation

Capallera Guirado, I. Identification and localization of septal flash by the means of time distributed networks and class activation maps. 2022. handle: http://hdl.handle.net/10230/54613

This citation was generated automatically.

Abstract

Cardiac resynchronization therapy (CRT) has been a significant step forward in the treatment of patients with arrhythmias leading to heart failure. However, it still presents two main issues. Firstly, the criteria to be a candidate for CRT are too simplistic, based on basic indices. Second, the therapy still has a 30% failure rate. Several studies have concluded that specific myocardial motion patterns, mainly the septal flash (SF), are related to a higher success rate of CRT. In prior works iden- tifying SF in left bundle branch block (LBBB) patients, a dyssynchrony abnormal pattern was present in 45% of the total patients. Its proper identification could serve both as a predictor of CRT success and as a more accurate condition to look for in CRT candidates, since SF is being detected manually by experienced clinicians these days, leading to high inter-observer variability. Most approaches to distinguish abnormal conditions rely on medical data. There- fore, machine learning and, concretely, neural networks can be key to revealing new patterns of deformation and improving the characterization and quantification of myocardial motion problems. These facts make SF identification a great subject of study for neural networks, thus potentially impacting clinical practice. However, septal flash occurs at a very specific point of time-space during the cardiac cycle. While neural networks have been demonstrated to be extremely accurate at finding such patterns, they have a huge lack of interpretability, a mandatory feature for clin- ical applications. In this study, I propose a method to analyze 2D ultrasound video sequences of cardiac motion and automatically characterize septal flash in an inter- pretable way. This is achieved by defining a time-distributed neural network with an attention system based on gradient class activation maps. The system identifies the most critical frames on the sequence where SF appears and marks the spatial region where the deformation takes place. The final project had 300 available sequences, 241 train and 59 test, and achieved a maximum accuracy of 0.80.