540 Assessment of left ventricular ejection fraction from echocardiographic images using machine learning algorithm

Background

Precise and reliable echocardiographic assessment of LVEF is needed for clinical decision-making. LVEF is currently determined through an observer dependent process that requires manual tracing. To remove this manual tracing step, which is both time-consuming and user dependent, automatic computer aided diagnosis systems may be useful in the clinical setting.

Purpose

The aim of this study was to evaluate whether a 3-dimensional convolutional neural networks (3DCNN) could estimate left ventricular ejection fraction (LVEF) and differentiate types of heart failure (preserved EF/reduced EF) using conventional 2-dimensional echocardiographic images.

Methods

We developed a deep learning model to automatically estimate LVEF from echocardiographic data. The 3DCNN model was trained on a dataset of 340 patients. The dataset creation consisted of three main steps: firstly, for each exam, cine-loops showing the parasternal and apical views were manually selected; then, 10 sequential frames were extracted from each 1 beat and; finally, each frame was pre-processed to fit the learning model. Each patient has 2 views, resulting in a total of 6,800 images. Reference LVEF measurement was calculated by two highly experienced readers in each case.

Results

A good correlation was found between estimated LVEF based on apical 2 and 4 chamber views and reference LVEF (r =0.88, p <0.001) (Figure). For classification of heart failure types based on LVEF (LVEF ≥50% or <50%), the area under the receiver-operating characteristic curve by the 3DCNN algorithm was over 0.95.

Conclusions

The 3DCNN can be applied to estimate and classify the LVEF in the clinical setting. Furthermore, this work will serve as a driver for future research using million image databases.

Abstract 540 Figure.

SU-E-J-203: Determination of PTV Margin for Lung Tumor Using In-Treatment 4D CBCT

PURPOSE

To determine a planning target volume (PTV) margin for lung cancer patients using a four-dimensional cone-beam CT (4D CBCT) acquired during volumetric modulated arc therapy (VMAT) treatment.

METHODS

A VMAT plan for lung cancer patients was created by Pinnacle v9.0 (Philips) treatment planning system (TPS), where the gross target volume (GTVs) in each breathing phase was delineated by using 4D-planning CT scan (TOSHIBA and ANZAI). The VMAT treatment was performed with a stereotactic body frame after the registration using Elekta X-ray volume imaging (XVI) unit. Simultaneous cone-beam projection images were acquired for 3 or 4 fractions of 10 patients. The in-treatment 4D CBCT was reconstructed by dividing into four breathing phase bins. A total of 38 in-treatment 4D-CBCT sets were exported to Pinnacle TPS. The isocenter of in-treatment 4D CBCT was matched with that of 4D-planning CT. The tumor motion during treatment was manually tracked on in-treatment 4D CBCT, and the center-of-mass (COM) location of the tumor was estimated. Analyzing the tumor regions observed by in-treatment 4D CBCT, a PTV margin in our system was derived.

RESULTS

The average difference in COM location of the tumor was less than 1mm for all directions, while the standard deviations (SD's) were about 1.3mm, 1.6mm, and 2.1mm for the lateral, the vertical, and the longitudinal directions, respectively. The large discrepancy more than 3mm was observed for one patient. The required PTV margin was about 3-4mm for the lateral and the vertical directions, whereas it was about 5mm for the longitudinal direction.

CONCLUSIONS

The uncertainties of the tumor motion caused by respiration were observed by in-treatment 4D CBCT images. It was feasible to determine the PTV margin from 4D volume images. K. Nakagawa receives research funding from Elekta.