/ChallengeHC18

Data Science Competition: “Automated Measurement of Fetal Head Circumference”. Top 8% Finalist.

Primary LanguageJupyter Notebook

ChallengeHC18

Automated measurement of fetal head circumference

Challenge goals

During pregnancy, ultrasound imaging is used to measure fetal biometrics. One of these measurements is the fetal head circumference (HC).

The HC can be used to estimate the gestational age and monitor growth of the fetus. The HC is measured in a specific cross section of the fetal head, which is called the standard plane.

Illustration

This challenge makes it possible to compare developed algorithms for automated measurement of fetal head circumference in 2D ultrasound images.

Data description

The dataset for this challenge contains a total of 1334 two-dimensional (2D) ultrasound images of the standard plane that can be used to measure the HC.

The data is divided into a training set of 999 images and a test set of 335 images. The size of each 2D ultrasound image is 800 by 540 pixels with a pixel size ranging from 0.052 to 0.326 mm.

The training set also includes an image with the manual annotation of the head circumference for each HC, which was made by a trained sonographer.

Metric

The results should be submitted as a csv file which contains 6 columns and 336 rows. The first row should be:

filename,center_x_mm,center_y_mm,semi_axes_a_mm,semi_axes_b_mm,angle_rad

Illustration

Score updates

Date Model LB score
Mean abs difference (mm) ± std		 Rank (TOP %) Solution weight_name
21/07 First commit x x x x
25/07 UNet
Batchnorm		 5.99 ± 11.26 554/848 (65%) SGD
5 epochs
no lr scheduling
dice loss		 UNet1_loss=0.63_SGD_ep=5_(216, 320)
26/07 UNet
Batchnorm		 4.26 ± 7.58 511/850 (60%) SGD
25 epochs
multistepLR
dice loss		 UNet1_dice=0.4405_SGD_ep=29_(216, 320)_wd=0_dice_loss
27/07 UNet
Batchnorm		 3.97 ± 6.97 503/851 (59%) SGD
35 epochs
multistepLR
dice loss
data aug		 UNet1_dice=0.438_SGD_ep=23_(216, 320)_wd=0_dice_loss
27/07 DilatedUNet
Batchnorm		 2.65 ± 6.08 402/852 (47%) SGD
35 epochs
multistepLR
dice loss
data aug		 DilatedUNet_dice=0.477_SGD_ep=28_(216, 320)_wd=0_dice_loss
27/07 DilatedUNet
Batchnorm		 2.42 ± 4.63 373/853 (43%) SGD
35 epochs
multistepLR
dice + WBCE loss
data aug		 DilatedUNet_dice=0.477_SGD_ep=24_(216, 320)_wd=0_bce_dice_loss
29/07 DilatedUNet
Batchnorm		 2.28 ± 3.53 351/853 (41%) SGD
35 epochs
multistepLR
dice + WBCE loss
more data aug
weight decay		 DilatedUNet_dice=0.478_SGD_ep=28_(216, 320)_wd=0.0001_bce_dice_loss
01/08 Custom UNet
Dilation
Deep supervision		 1.92 ± 2.33 73/864 (8%) SGD
40 epochs
multistepLR
dice + WBCE loss
data aug
weight decay		 U_Net_dice=0.476_SGD_ep=30_(216, 320)_wd=0.0001_bce_dice_loss
03/08 Custom UNet
Dilation
Deep supervision		 1.91 ± 1.89 72/876 (8%) SGD
60 epochs
cosine annealing
dice + WBCE loss
data aug
weight decay		 U_Net_dice=0.483_SGD_ep=51_(216, 320)_wd=0.0001_bce_dice_loss_cosineannealing

Final solution

Example

  • Data preprocessing:

    • train/test/valid (80%/15%/5%)

  • Architecture: unet + dilation in bottleneck + deep supervision

  • Ellipse fitting: https://ieeexplore.ieee.org/document/546029/

  • Training:

    • data aug: h/v flip, scale, shift
    • resize to (216, 320)
    • random weights
    • SGD
    • LR scheduler : MultiStepLR
    • 1e-4 weight decay
    • Loss : BCE + DICE

What didn't work

  • Focal loss
  • Tversky loss
  • Adam/AdamW (converges fast but not as good as SGD)
  • Nested UNet (UNet++)
  • Attention Unet
  • Multi input

Things I would have liked to try

  • TTA
  • Other architectures (maskrcnn, deeplabv3+, etc.)
  • different input sizes
  • weights from pretrained autoencoder
  • other algorithms for ellipse fitting
  • ensemble

Fails

Example