A cardiac function examination conducted by a trained radiologist from cardiac cycles is predisposed to notable inter-observer variability. The segmentation of Echocardiography (ECHO), a cardiac ultrasound image, is critical due to its low contrast, poor quality imaging, and a 2D replica of a 3D continuously moving organ, introducing many sources of variability. For cardiac function analysis, the left ventricle (LV) segmentation on both end-diastole (ED) and end-systole (ES) frames is required, and the ES frame necessitates additional attention during manual LV delineation due to the heart’s contraction and complexity compared to the ED frame. To address these problems, an automatic deep learning architecture called Parallel Left Ventricle Segmentation Network (PLVS-Net) is proposed. The PLVS-Net comprises two parallel networks (modified U-Nets) that accurately delineate the LV during ED and ES phases simultaneously and later concatenate the results to measure clinical indices. The frames corresponding to ED and ES phases are extracted from echocardiography videos in the EchoNet-Dynamic dataset. The PLVS-Net outperformed state-of-the-art segmentation methods on the EchoNet dataset, achieving a Dice similarity coefficient (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42979_2025_4155_Article_IEq1.gif" Format="GIF" Height="19" Rendition="HTML" Resolution="72" Type="Linedraw" Width="66" /> </InlineMediaObject> <EquationSource Format="TEX">$DSC (\%)$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>D</mi> <mi>S</mi> <mi>C</mi> <mo stretchy="false">(</mo> <mo>%</mo> <mo stretchy="false">)</mo> </mrow> </math></EquationSource> </InlineEquation>) of <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42979_2025_4155_Article_IEq2.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="87" /> </InlineMediaObject> <EquationSource Format="TEX">$93.44 \pm 1.18$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>93.44</mn> <mo>±</mo> <mn>1.18</mn> </mrow> </math></EquationSource> </InlineEquation>, a mean absolute distance (MAD) of <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42979_2025_4155_Article_IEq3.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="80" /> </InlineMediaObject> <EquationSource Format="TEX">$0.94 \pm 0.37$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0.94</mn> <mo>±</mo> <mn>0.37</mn> </mrow> </math></EquationSource> </InlineEquation>, and a Hausdorff distance (HD) of <InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42979_2025_4155_Article_IEq4.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="80" /> </InlineMediaObject> <EquationSource Format="TEX">$3.23 \pm 2.43$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>3.23</mn> <mo>±</mo> <mn>2.43</mn> </mrow> </math></EquationSource> </InlineEquation>. Additionally, without fine-tuning a DSC <InlineEquation ID="IEq5"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42979_2025_4155_Article_IEq5.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="101" /> </InlineMediaObject> <EquationSource Format="TEX">$91.53\% \pm 4.80$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>91.53</mn> <mo>%</mo> <mo>±</mo> <mn>4.80</mn> </mrow> </math></EquationSource> </InlineEquation>, and after fine-tuning on the CAMUS dataset, PLVS-Net delivered even better results. The segmented ED and ES frames of the left ventricle are utilized to calculate clinical indices, precisely the left ventricular end-diastolic volume (<InlineEquation ID="IEq6"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42979_2025_4155_Article_IEq6.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="51" /> </InlineMediaObject> <EquationSource Format="TEX">$LV_{EDV}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>L</mi> <msub> <mi>V</mi> <mrow> <mi mathvariant="italic">EDV</mi> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation>) and left ventricular end-systolic volume (<InlineEquation ID="IEq7"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42979_2025_4155_Article_IEq7.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="49" /> </InlineMediaObject> <EquationSource Format="TEX">$LV_{ESV}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>L</mi> <msub> <mi>V</mi> <mrow> <mi mathvariant="italic">ESV</mi> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation>), using Simpson’s biplane method. These volumes, <InlineEquation ID="IEq8"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42979_2025_4155_Article_IEq6.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="51" /> </InlineMediaObject> <EquationSource Format="TEX">$LV_{EDV}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>L</mi> <msub> <mi>V</mi> <mrow> <mi mathvariant="italic">EDV</mi> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation> and <InlineEquation ID="IEq9"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42979_2025_4155_Article_IEq7.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="49" /> </InlineMediaObject> <EquationSource Format="TEX">$LV_{ESV}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>L</mi> <msub> <mi>V</mi> <mrow> <mi mathvariant="italic">ESV</mi> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation>, are then used to compute the left ventricular ejection fraction (<InlineEquation ID="IEq10"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42979_2025_4155_Article_IEq10.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="42" /> </InlineMediaObject> <EquationSource Format="TEX">$LV_{EF}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>L</mi> <msub> <mi>V</mi> <mrow> <mi mathvariant="italic">EF</mi> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation>). The mean squared error for predicting <InlineEquation ID="IEq11"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="42979_2025_4155_Article_IEq10.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="42" /> </InlineMediaObject> <EquationSource Format="TEX">$LV_{EF}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>L</mi> <msub> <mi>V</mi> <mrow> <mi mathvariant="italic">EF</mi> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation> is 4.01 on the EchoNet dataset. PLVS-Net has demonstrated superior performance compared to previously reported networks, with reduced throughput delay, making it well-suited for real-time applications when ED and ES frames are available.

PLVS-Net: a Parallel Left Ventricle Segmentation Network for Clinical Indices Measurement in 2D Echocardiography

期刊论文
发表日期：

2025年7月10日

查看全文

Gajraj Singh,
Anand D. Darji,
Jignesh N. Sarvaiya,
Suprava Patnaik

摘要

A cardiac function examination conducted by a trained radiologist from cardiac cycles is predisposed to notable inter-observer variability. The segmentation of Echocardiography (ECHO), a cardiac ultrasound image, is critical due to its low contrast, poor quality imaging, and a 2D replica of a 3D continuously moving organ, introducing many sources of variability. For cardiac function analysis, the left ventricle (LV) segmentation on both end-diastole (ED) and end-systole (ES) frames is required, and the ES frame necessitates additional attention during manual LV delineation due to the heart’s contraction and complexity compared to the ED frame. To address these problems, an automatic deep learning architecture called Parallel Left Ventricle Segmentation Network (PLVS-Net) is proposed. The PLVS-Net comprises two parallel networks (modified U-Nets) that accurately delineate the LV during ED and ES phases simultaneously and later concatenate the results to measure clinical indices. The frames corresponding to ED and ES phases are extracted from echocardiography videos in the EchoNet-Dynamic dataset. The PLVS-Net outperformed state-of-the-art segmentation methods on the EchoNet dataset, achieving a Dice similarity coefficient ( $DSC (\%)$ $D S C (%)$ ) of $93.44 \pm 1.18$ $93.44 \pm 1.18$ , a mean absolute distance (MAD) of $0.94 \pm 0.37$ $0.94 \pm 0.37$ , and a Hausdorff distance (HD) of $3.23 \pm 2.43$ $3.23 \pm 2.43$ . Additionally, without fine-tuning a DSC $91.53\% \pm 4.80$ $91.53 % \pm 4.80$ , and after fine-tuning on the CAMUS dataset, PLVS-Net delivered even better results. The segmented ED and ES frames of the left ventricle are utilized to calculate clinical indices, precisely the left ventricular end-diastolic volume ( $LV_{EDV}$ $L V_{EDV}$ ) and left ventricular end-systolic volume ( $LV_{ESV}$ $L V_{ESV}$ ), using Simpson’s biplane method. These volumes, $LV_{EDV}$ $L V_{EDV}$ and $LV_{ESV}$ $L V_{ESV}$ , are then used to compute the left ventricular ejection fraction ( $LV_{EF}$ $L V_{EF}$ ). The mean squared error for predicting $LV_{EF}$ $L V_{EF}$ is 4.01 on the EchoNet dataset. PLVS-Net has demonstrated superior performance compared to previously reported networks, with reduced throughput delay, making it well-suited for real-time applications when ED and ES frames are available.