Automatic measurement of fetal anterior neck lower jaw angle in nuchal translucency scans

Abstract

This study aims at suggesting an end-to-end algorithm based on a U-net-optimized generative adversarial network to predict anterior neck lower jaw angles (ANLJA), which are employed to define fetal head posture (FHP) during nuchal translucency (NT) measurement. We prospectively collected 720 FHP images (half hyperextension and half normal posture) and regarded manual measurement as the gold standard. Seventy percent of the FHP images (half hyperextension and half normal posture) were used to fit models, and the rest to evaluate them in the hyperextension group, normal posture group (NPG), and total group. The root mean square error, explained variation, and mean absolute percentage error (MAPE) were utilized for the validity assessment; the two-sample t test, Mann-Whitney U test, Wilcoxon signed-rank test, Bland-Altman plot, and intraclass correlation coefficient (ICC) for the reliability evaluation. Our suggested algorithm outperformed all the competitors in all groups and indices regarding validity, except for the MAPE, where the Inception-v3 surpassed ours in the NPG. The two-sample t test and Mann-Whitney U test indicated no significant difference between the suggested method and the gold standard in group-level comparison. The Wilcoxon signed-rank test revealed significant differences between our new approach and the gold standard in personal-level comparison. All points in Bland-Altman plots fell between the upper and lower limits of agreement. The inter-ICCs of ultrasonographers, our proposed algorithm, and its opponents were graded good reliability, good or moderate reliability, and moderate or poor reliability, respectively. Our proposed approach surpasses the competition and is as reliable as manual measurement.

Figure 1

Data collection and preprocessing. (a) Original image of anterior neck lower jaw angle (ANLJA); (b) illustration of ANLJA measurement; (c) raw ANLJA image from an ultrasonoscope; (d) 400 × 400-pixel screenshot of ANLJA captured manually.

Figure 2

Comparison of different ANLJA prediction algorithms in validity assessment. (a), (b), and (c) are the bar charts of root mean square error (RMSE), explained variation (EVA), and mean absolute percentage error (MAPE), respectively. The smaller the RMSEs and MAPEs, the higher the performance; the bigger the EVA, the better the models operate. TG, NPG, and HG denote the Total group, Normal posture group, and Hyperextension group, respectively.

Figure 3

Comparison of algorithms and ultrasonographers in reliability assessment. (a) Group-level comparison of mean values of suggested method and manual measurement with two-sample t test (standard deviation bars). (b) Group-level comparison of medians of suggested method and manual measurement with Mann–Whitney U test and illustration of distributions of differentials between suggested method and manual measurement; (c) and (d) Personal-level comparison of suggested method to manual measurement with Wilcoxon signed-rank test. (e) and (f) Bland–Altman plots for reliability assessment between proposed algorithm and manual measurement. (g) Reliability evaluation for algorithms and ultrasonographers by intraclass correlation coefficient (ICC) with 95% confidence intervals. Poor reliability: ICCs below 0.50; moderate reliability: ICCs from 0.50 to 0.75; good reliability: ICCs between 0.75 and 0.90; excellent reliability: ICCs above 0.90. Two ultrasonographers’ ICCs are intra-ICCs, and the rest ICCs are inter-ICCs. NPG, Normal posture group; HG, Hyperextension group; TG, Total group; LoA, limit of agreement.

Figure 4

Flowchart of inclusion, exclusion, and quality control sampling. HPMCHCH, Hunan Provincial Maternal and Child Health Care Hospital; QCG, quality control group; NQCG, non-quality control group; NT, nuchal translucency; ICC, intraclass correlation coefficient.

Figure 5

Structure of proposed ANLJA prediction algorithm. (a) Structure of U-net-based generator network; (b) Structure of discriminator network derived from Wasserstein generative adversarial network with a gradient penalty; (c) Structure of multi-level multi-scale receptive field residual modular block (MMRFRMB); (d) Structure of receptive field residual modular block (RFRMB); (e) Structure of receptive field residual module (RFRM); (f) Structure of receptive field dense block (RFDB); (g) Structure of receptive field block with small cores (RFBs). LSC, long skip connection; RAM denotes hybrid attention mechanism; α stands for feature weight.