Development and validation of an accurate smartphone application for measuring waist-to-hip circumference ratio

Waist-to-hip circumference ratio (WHR) is now recognized as among the strongest shape biometrics linked with health outcomes, although use of this phenotypic marker remains limited due to the inaccuracies in and inconvenient nature of flexible tape measurements when made in clinical and home settings. Here we report that accurate and reliable WHR estimation in adults is possible with a smartphone application based on novel computer vision algorithms. The developed application runs a convolutional neural network model referred to as MeasureNet that predicts a person’s body circumferences and WHR using front, side, and back color images. MeasureNet bridges the gap between measurements conducted by trained professionals in clinical environments, which can be inconvenient, and self-measurements performed by users at home, which can be unreliable. MeasureNet’s accuracy and reliability is evaluated using 1200 participants, measured by a trained staff member. The developed smartphone application, which is a part of Amazon Halo, is a major advance in digital anthropometry, filling a long-existing gap in convenient, accurate WHR measurement capabilities.


Neck
• Participant standing in relaxed posture with head in neutral position.
• Locate the upper end of the neck where the neck meets the jaw • Locate the base of the neck, just above the shoulder.
• Measure is taken at midway point between two above locations.Measuring tape is perpendicular to neck axis.

Chest For Men / Bust for Women
• Participant standing in relaxed A-pose (best effort for self-measure).
• Measure is taken at area of maximal circumference of upper torso region.
• Take care to ensure that tape measure does not have excess distance across sides and back.

Arm
• Measures to be taken of the right arms.
• Participant standing in relaxed posture with arm to be measured at a 90 degree angle with the palm facing up.• Follow the participant's spine of the right scapula until it makes a sharp V at the shoulder.Note the shoulder at this location.
• Measure the arm from this point to the tip of the elbow and carefully make sure the measuring tape is in the center of the posterior surface of the arm.Note the midpoint of the arm.• Ask the participant to stand with both arms hanging loosely at their side and weight evenly distributed on both feet.The participant should not flex or tighten any muscles.• Measure the circumference at the previous noted midpoint of the arm.The tape should be perpendicular to the long axis of the upper arm.

Waist
• Have participant stand with weight evenly distributed on both feet in relaxed A-pose.• Locate the right ilium of the pelvis and draw a line just above the lateral border of the right ilium.• Place the measuring tape around the participant making sure it is snug but not compressing skin.The measuring tape should be completely parallel to the ground.

Hip
• Have participant stand with weight evenly distributed on both feet.• The measuring tape is placed around the buttocks at the point of maximal circumference.The sides of the tape should be checked to ensure the tape is horizontal.The tape should be held snug but not tight.• The value should be read at the right side of the participant.

Thigh
• Measures to be taken of the right thigh.
• Locate midpoint of thigh by placing zero end of measuring tape at the inguinal crease.Folds of fat tissue may have to be lifted to reach the crease.Extend tape down towards the line created at the proximal end of the patella.To be sure you have the zero end of the tape at the inguinal crease, place your thumb at the zero end of the tape and instruct the participant to slightly lift the thigh.A tightening of muscle tendon should be felt.Note the midpoint.• Measure from the patella to 1/3 the length.Note this point.
• After noting the midpoint, have the participant stand with most of his or her weight on the non-measured leg.The measured leg should be slightly bent and forward.The participant may use an object (such as a table) for balance.• Place the measuring tape around the mid-thigh as noted.The tape should be perpendicular to the long axis of the thigh.

Calf
• Measures to be taken of the right calf.
• Locate the point of maximal circumference between the knee and ankle.
• Have the participant stand in a relaxed A-pose.
• Wrap the measuring tape around the lower leg at the noted point of maximal circumference.
• The tape should be perpendicular to the long axis of the lower leg.

Forearm
• Measures to be taken of the right forearm.
• Have the participant stand with arm hanging loosely at their side and weight evenly distributed on both feet.The participant should not flex or tighten any muscles.
• Locate the point of maximal circumference between the elbow and wrist.
• Wrap the measuring tape around the forearm at the noted point of maximal circumference.
• The tape should be perpendicular to the long axis of the forearm.

Synthetic Ground Truth Measurement
The method to estimate ground truth measurement rings on the synthetic mesh that align with the tape measured ground truth was as follows.First, 112 dense circumferences were defined over the body as shown in

lower right leg circumferences
Each circumference was found by intersecting a plane with the SMPL mesh at specific predefined intersection points and finding the circumference at that intersection.These points and the corresponding intersection planes were found as follows: • We first find 23 3D joint locations over the mesh using the SMPL parametric joint regressor.We also define skeletal axes that connect these 3D joints (as shown by green lines in Figure 2).

•
We sub-divide each axis based on the average body part length to find the intersection points.For e.g., upper left/right arm circumferences consist of 8 locations where intersections are found between the left shoulder joint and left elbow joint locations respectively.To make sure that the plane does not selfintersect at multiple points on the mesh, we segment the part specific vertices and faces before finding the intersection.
• Lastly, we refine the orientation of the plane before intersection.We reestimate the orientation such that intersecting plane is perpendicular to the surface of the mesh for the specified body part.This ensures a realistic and robust circumference estimation method that can generalize across the variation of body shapes.
Next, we identify the correct part specific circumference locations on the SMPL mesh which aligns with tape measurements.We find these locations using a data-driven approach using a dataset containing 3D scans and tape measurements.For each part, we collect dense measurements using the above protocol and compare the circumference with the mean of two tape measurements taken by a trained staff member.The location index that minimizes the MAE (Eq. 1) for each part is considered as ground truth on SMPL meshes.This reduces the output domain gap between the ground truth circumference definition on synthetic meshes which used to train The table below compares the mean absolute error in hip and waist as compared to staff tape measurement (from CSD dataset) given the MeasureNet predictions at and around optimal ring location.Accuracy is relative to ground truth staff measurements on Circumference Study Dataset (CSD).We compare the error between optimal ring index and two rings around the optimal index.It shows that the predicted error at optimal ring index is the lowest when compared to ring indices around it.

Predicting Waist-Hip ratio through Regression vs. Classification
WHR can be predicted indirectly (by taking ratios of waist and hip estimates) and directly (either through regression or classification).The table below compares the accuracy using different methods.We use the average of WHR predictions through classifications, regression and ratio of predicted waist and hip circumferences as the final WHR prediction.We found that the average prediction was the most robust without losing accuracy as compared to predictions via regression, classification or taking the ratio of waist and hip.Accuracy is relative to ground truth staff measurements on Circumference Study Dataset (CSD).

Figure 1 .
The dense circumferences (shown in red) consisted of the following: MeasureNet and tape measurements taken by a trained staff member.Ground truth circumference rings which align with tape measurements are shown in Figure 2 for the segmented regions.Supplementary Figure 1.Dense circumference locations.The MeasureNet algorithm predicts 112 circumferences defined densely over the body.Dense circumferences are shown in red and are uniformly sampled over the body.Supplementary Figure 2. Tape measurements aligned synthetic circumference rings.Segmented vertices and faces are shown in color.3D joint locations are shown in pink.Ground truth circumference rings which align with tape measurements are shown in white for the segmented regions.

of-the Art Measurements Men Women n= 73 83 Accuracy & Noise Repeatability State-of-the Art Comparisons n= 4. Additional ablation results with ReLU activations, Self-attention (SA) and Squeeze-Excitation (SE) blocks.
Disposition of Study Participants.The table below compares MeasureNet models trained with ReLU or Swish activation, with or without Self-attention (SA) and Squeeze-Excitation (SE) blocks.The models were tested on Human Solutions dataset.Lower is better.

MeasureNet with uncertainty-based loss weighting 4
Since we have multiple loss functions corresponding to each output in MeasureNet, hand-tuning each loss weight is expensive and fragile.Based on Kendall et al.4, we used uncertainty-based loss weighting where the weight parameter is learned.The table below shows the improvement in accuracy using uncertainty-based loss weighting when tested on Human Solutions dataset.Lower is better.

MeasureNet using synthetic textured color images
The table shows that training a model with synthetic textured color images generalizes poorly when tested on real examples from Human Solutions dataset as compared to using 23-class segmented images as input.These models were trained using a subset of synthetic dataset.The numbers in the table show the circumference prediction error when using color image or 23-class segmented image as an input.Lower is better. .Demographic and anthropometric characteristics for Circumference Study Dataset.BMI, body mass index.Data are X± 3SD. .Demographic and anthropometric characteristics for Human Solutions dataset used in comparison with state-of-the-art approaches.Data are X± 3SD. AB