Table 3 Dice and mIoU scores of various ablations on the MyoVision-US AI model on the holdout validation set.

From: Development of an artificial intelligence powered software for automated analysis of skeletal muscle ultrasonography

Epochs

Strong color augmentations

Backbone architecture

Decoder architecture

Parameters (Millions)

Loss function

QC Dice

QC mIoU

TA Dice

TA mIoU

40

Yes

ResNet-50

DeepLabV3

39.6

BCE

0.900

0.881

0.949

0.904

40

Yes

ResNet-50

UNet

32.5

BCE

0.893

0.875

0.944

0.895

40

Yes

MIT-B2

SegFormer

24.7

BCE

0.883

0.866

0.956

0.916

40

Yes

MIT-B3

SegFormer

44.6

BCE

0.880

0.860

0.955

0.914

20

Yes

ResNet-50*

DeepLabV3*

39.6*

BCE

0.881

0.861

0.946

0.898

80

Yes

ResNet-50*

DeepLabV3*

39.6*

BCE

0.884

0.865

0.943

0.893

40

No

ResNet-50*

DeepLabV3*

39.6*

BCE

0.897

0.876

0.950

0.905

40

Yes

ResNet-50*

DeepLabV3*

39.6*

Focal

0.883

0.866

0.935

0.878

40

Yes

ResNet-50*

DeepLabV3*

39.6*

Tversky

0.851

0.830

0.942

0.892

40

Yes

ResNet-50*

DeepLabV3*

39.6*

Jaccard

0.867

0.849

0.950

0.905

40

Yes

ResNet-50*

DeepLabV3*

39.6*

Lovasz

0.846

0.824

0.953

0.910

40

Yes

ResNet-50*

DeepLabV3*

39.6*

Dice

0.891

0.873

0.950

0.905

  1. QC = quadriceps complex (mean of specific measure, e.g. Dice, over rectus femoris, vastus intermedius, and femur). TA = tibialis anterior. mIoU = mean Intersection over Union. BCE = binary cross-entropy.
  2. *Indicates the backbone architecture and decoder architecture used for TA ablations were MIT-B2 and SegFormer while QC ablations were performed with ResNet-50 and DeepLabV3.
Back to article page