Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ConFormer: A Novel Collection of Deep Learning Models to Assist Cardiologists in the Assessment of Cardiac Function

Ethan Thomas, Salman Aslam

TL;DR

Confronting manual, observer-variant measurement of cardiac function from echocardiograms, the paper introduces ConFormer, a compact DL framework that decouples spatial, channel, and temporal processing for efficient LVEF and LV dimension estimation. The LVEF pathway uses a Beat Extractor based on DeepLabV3+ segmentation and a lightweight Xception-based Predictor with Kronecker-factorized temporal convolutions, achieving a MAE of $6.57$ on EchoNet-Dynamic with only $5.82$M parameters. The LVD pathway applies a DeepLabV3+-based landmark detector with EfficientNetV2S to measure IVS, LVID, and LVPW, obtaining a combined MAE around $6$ and competitive performance versus state-of-the-art models across resolutions. Overall, ConFormer demonstrates accurate, scalable, and resource-efficient automated cardiac measurements, with code publicly available for validation and deployment ($EF = \frac{EDV - ESV}{EDV}$ in standard EF computation).

Abstract

Cardiovascular diseases, particularly heart failure, are a leading cause of death globally. The early detection of heart failure through routine echocardiogram screenings is often impeded by the high cost and labor-intensive nature of these procedures, a barrier that can mean the difference between life and death. This paper presents ConFormer, a novel deep learning model designed to automate the estimation of Ejection Fraction (EF) and Left Ventricular Wall Thickness from echocardiograms. The implementation of ConFormer has the potential to enhance preventative cardiology by enabling cost-effective, accessible, and comprehensive heart health monitoring, thereby saving countless lives. The source code is available at https://github.com/Aether111/ConFormer.

ConFormer: A Novel Collection of Deep Learning Models to Assist Cardiologists in the Assessment of Cardiac Function

TL;DR

Confronting manual, observer-variant measurement of cardiac function from echocardiograms, the paper introduces ConFormer, a compact DL framework that decouples spatial, channel, and temporal processing for efficient LVEF and LV dimension estimation. The LVEF pathway uses a Beat Extractor based on DeepLabV3+ segmentation and a lightweight Xception-based Predictor with Kronecker-factorized temporal convolutions, achieving a MAE of on EchoNet-Dynamic with only M parameters. The LVD pathway applies a DeepLabV3+-based landmark detector with EfficientNetV2S to measure IVS, LVID, and LVPW, obtaining a combined MAE around and competitive performance versus state-of-the-art models across resolutions. Overall, ConFormer demonstrates accurate, scalable, and resource-efficient automated cardiac measurements, with code publicly available for validation and deployment ( in standard EF computation).

Abstract

Cardiovascular diseases, particularly heart failure, are a leading cause of death globally. The early detection of heart failure through routine echocardiogram screenings is often impeded by the high cost and labor-intensive nature of these procedures, a barrier that can mean the difference between life and death. This paper presents ConFormer, a novel deep learning model designed to automate the estimation of Ejection Fraction (EF) and Left Ventricular Wall Thickness from echocardiograms. The implementation of ConFormer has the potential to enhance preventative cardiology by enabling cost-effective, accessible, and comprehensive heart health monitoring, thereby saving countless lives. The source code is available at https://github.com/Aether111/ConFormer.
Paper Structure (19 sections, 5 equations, 4 figures, 1 table)

This paper contains 19 sections, 5 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Previous Work
  3. Datasets
  4. LVEF Estimation
  5. LVD Estimation
  6. Video Analysis Using CNNs
  7. LVEF (Left Ventricular Ejection Fraction) Estimation Model
  8. Data
  9. Design
  10. Decoupling the Cross-Channel Convolutions
  11. Decoupling the Temporal Convolutions
  12. Beat Extractor
  13. Prediction Module
  14. Results and Discussion
  15. LVD (Left Ventricular Dimensions) Estimation Model
  16. ...and 4 more sections

Figures (4)

  • Figure 1: The LVEF prediction pipeline consists of heart-beat extraction, spatial feature extraction, temporal analysis and output regression (top). The heart-beat extractor is based on the DeepLabV3+ architecture (bottom).
  • Figure 2: Segmentation of left ventricular cardiac chamber during diastole (left figures) and systole (rights figures) during the same heartbeat (above). Predicted vs manually calculated ejection fraction (below).
  • Figure 3: The LVD (left ventricular dimensions) prediction pipeline.
  • Figure 4: LV dimensions.