Validation of deep learning enabled software MetronMind to measure vertebral heart size and vertebral left atrial size in dogs

Figures

• 

  Fig 1. Example of radiographs with variable landmark quality. A) Radiograph with all landmarks classified as excellent. B) Radiograph with poor left ventricular apex and cranial waist; poor vena cava; excellent vertebral bodies and T4; and diagnostic carina.

• 

  Fig 2. Examples of traditional (A), modified (B) and MetronMind (C) measurements of vertebral heart size (VHS) and vertebral left atrial size (VLAS). A) yellow lines indicate traditional VHS measurements and blue lines indicate traditional VLAS. B) Red line indicates the measurement of five vertebral bodies to calculate modified VHS and VLAS, yellow line the left ventricular (LV) major axis, green line the LV minor axis for VHS and blue line VLAS. C) A green line indicates the measurement of five vertebral bodies, orange lines measures of VHS and blue line VLAS as generated by the MetronMind algorithm.

• 

  Fig 3. Comparison of traditional and modified methods of measuring vertebral heart size and vertebral left atrial size (VHS-T vs VHS-M and VLAS-T vs VLAS-M) on right lateral radiographs. A) When comparing VHS-T and VHS-M Passing–Bablok analysis yielded the equation VHS-M = 1.06(VHS-T)–0.58, Pearson’s linear correlation coefficient r = 0.994, n = 1055. Blue line: fitted regression line, shading: 95% confidence interval, red dashed line: identity. B) Bland–Altman analysis demonstrated a mean negative bias of −0.10 vertebrae (95% CI −0.11 to −0.10) for the VHS-M method. Dashed lines give limits of agreement (±1.96 SD) and mean bias; shading gives 95% confidence intervals with dotted lines at the limits. C) When comparing VLAS-T and VLAS-M Passing–Bablok analysis yielded the equation VLAS-M = 1.00(VLAS-T)–0.00, Pearson’s linear correlation coefficient r = 0.974, n = 1057. D) Bland–Altman analysis demonstrated a mean bias of 0.040 vertebrae (95% CI 0.032 to 0.046) for the modified VLAS (VLAS-M) method.

• 

  Fig 4. Comparison of modified and artificial intelligence (AI) algorithm (MetronMind) methods of measuring vertebral heart size and vertebral left atrial size (VHS-M vs VHS-AI and VLAS-M vs VLAS-AI) on right lateral radiographs. A) When comparing VHS-M and VHS-AI Passing–Bablok analysis yielded the equation VHS-AI = 0.95(VHS-M)+0.44, Pearson’s linear correlation coefficient r = 0.947, n = 1055. Blue line: fitted regression line, shading: 95% confidence interval, red dashed line: identity. B) Bland–Altman analysis demonstrated a mean bias of 0.08 vertebrae (95% CI 0.06 to 0.10) for the VHS-M compared to the VHS-AI method. Dashed lines give limits of agreement (±1.96 SD) and mean bias; shading gives 95% confidence intervals with dotted lines at the limits. C) When comparing VLAS-M and VLAS-AI Passing–Bablok analysis yielded the equation VLAS-AI = 0.83(VLAS-M)+0.28, Pearson’s linear correlation coefficient r = 0.811, n = 1055. D) Bland–Altman analysis demonstrated a mean bias of 0.07 vertebrae (95% CI 0.05 to 0.09) for VLAS-M compared to the VLAS-AI method.

• 

  Fig 5. Comparison of measuring modified vertebral heart size (VHS-M) and vertebral left atrial size (VLAS-M) on left (LL) and right (RL) lateral radiographs. A) When comparing measurement of VHS-M on left and right lateral radiographs Passing–Bablok analysis yielded the equation LL-VHS-M = 1.0(RL-VHS-M)-0.30, Pearson’s linear correlation coefficient r = 0.871, n = 445. Blue line: fitted regression line, shading: 95% confidence interval, red dashed line: identity. B) Bland–Altman analysis demonstrated a mean bias of 0.27 vertebrae (95% CI 0.21 to 0.32) for measurements made on the right compared to left radiographs. Dashed lines give limits of agreement (±1.96 SD) and mean bias; shading gives 95% confidence intervals with dotted lines at the limits. C) When VLAS-M on left and right lateral radiographs Passing–Bablok analysis yielded the equation LL-VLAS-M = 1.0(RL-VLAS-M)+0.28, Pearson’s linear correlation coefficient r = 0.726, n = 439. D) Bland–Altman analysis demonstrated a negative mean bias of −0.13 vertebrae (95% CI −0.16 to −0.10) for measurements made on the right compared to left radiographs.