Table 7 Validation results for three models from the DS and DNS pipelines on the CBIS-DDSM dataset, using input images of size 256 \(\times\) 256 with P1–P99 pixel intensity normalization, while varying resizing strategies (rectangular vs. square shapes, with or without padding).

From: The impact of pre-processing techniques on deep learning breast image segmentation

 

Model name

Padding

Keep aspect ratio

DSC ± STD

HD ± STD (in mm)

Comparison

p-value

A

DS 256 P1-P99

Yes

Yes

0.657 ± 0.269

27 ± 14

A

B

0.004

B

DNS 256 P1-P99

No

No

0.610 ± 0.229

40 ± 21

A

C

0.000

C

DNS 256 448 P1-P99

No

Yes

0.466 ± 0.205

85 ± 19

B

C

0.000

Terminology

Pipeline

Orientation

Pixel spacing

Pixel intensity

Resize

DS 256 P1-P99

DS

Yes

Yes

P1-P99

256 × 256

DNS 256 P1-P99

DNS

Yes

Yes

P1-P99

256 × 256

DNS 256 448 P1-P99

DNS

Yes

Yes

P1-P99

256 × 448

  1. The table reports the mean DSC and HD over 5-fold cross-validation for lesion segmentation, together with their standard deviations. The second part of the table provides a legend describing the model names.
  2. Bold values denote the best results.
Back to article page