Method, device, and storage medium for weakly-supervised universal lesion segmentation with regional level set loss

Patent metadata

Field	Value
Publication number	US-11900596-B2
Application number	US-202117479560-A
Country	US
Kind code	B2
Filing date	Sep 20, 2021
Priority date	Apr 14, 2021
Publication date	Feb 13, 2024
Grant date	Feb 13, 2024

Abstract

The present disclosure provides a computer-implemented method, a device, and a storage medium. The method includes inputting an image into an attention-enhanced high-resolution network (AHRNet) to extract feature maps for generating a first feature map; generating a first probability map which is concatenated with the first feature map to form a concatenated first feature map, and updating the AHRNet using the first segmentation loss; generating a second feature map, and scaling the second feature map to form a third feature map; generating a second probability map which is concatenated with the third feature map to form a concatenated third feature map, and updating the AHRNet using the second segmentation loss; generating a fourth feature map, and scaling the fourth feature map to form a fifth feature map; updating the AHRNet using the third segmentation loss and the regional level set loss; and outputting the third probability map.

First claim

What is claimed is:

1. A lesion segmentation method for medical images, comprising:
inputting an image into an attention-enhanced high-resolution network (AHRNet) to extract feature maps at multiple scales;
generating a first feature map according to the extracted feature maps;
generating a first probability map according to the first feature map, concatenating the first probability map with the first feature map to form a concatenated first feature map, calculating a first segmentation loss based on the first probability map, and updating the AHRNet using the first segmentation loss;
generating a second feature map by up-sampling the concatenated first feature map using a deconvolutional layer, and scaling the second feature map to form a third feature map;
generating a second probability map according to the third feature map, concatenating the second probability map with the third feature map to form a concatenated third feature map, calculating a second segmentation loss based on the second probability map, and updating the AHRNet using the second segmentation loss;
generating a fourth feature map by up-sampling the concatenated third feature map using a deconvolutional layer, and scaling the fourth feature map to form a fifth feature map;
generating a third probability map according to the fifth feature map, and calculating a third segmentation loss and a regional level set loss based on the third probability map; and
updating the AHRNet using the third segmentation loss and the regional level set loss, and outputting the third probability map.

2. The method according to claim 1 , wherein generating the first feature map according to the extracted feature maps includes:
inputting each feature map of the extracted feature maps at each scale into a dual attention (DA) module of multiple DA modules, thereby forming an enhanced feature map at each scale of multiple scales:
up-sampling one or more enhanced feature maps corresponding to one or more scales among the multiple scales, such that enhanced feature maps, including the up-sampled enhanced feature maps and non-up-sampled enhanced feature maps, have a same resolution;
inputting the enhanced feature maps having the same resolution into a scale attention (SA) module to generate concatenated feature maps; and
generating the first feature map according to the concatenated feature maps.

3. The method according to claim 1 , wherein:
the first segmentation loss is calculated according to the first feature map and a corresponding pseudo mask; the second segmentation loss is calculated according to the second feature map and a corresponding pseudo mask; and the third segmentation loss is calculated according to the third feature map and a corresponding pseudo mask, wherein the corresponding pseudo mask is an ellipse for the image, fitted from four endpoints of a response evaluation criteria in solid tumors (RECIST) annotation of the image.

4. The method according to claim 1 , wherein the regional level set loss is computed by:

ℓ
rls

=

1

❘
"\[LeftBracketingBar]"

I
′

❘
"\[RightBracketingBar]"

⁢

∑

i
∈

I
′

[

λ
1

·

p
⁡
(
i
)

·

❘
"\[LeftBracketingBar]"

q
⁡
(
i
)

-

c
1

❘
"\[RightBracketingBar]"

2

+

λ
2

·

(

1
-

p
⁡
(
i
)

)

·

❘
"\[LeftBracketingBar]"

q
⁡
(
i
)

-

c
2

❘
"\[RightBracketingBar]"

2

]

wherein λ 1 and λ 2 denote predefined non-negative hyper-parameters, q(i) is an intensity of a pixel i, c 1 and c 2 denote mean pixel intensities of inside and outside areas of a contour, p(i) denotes a probability value of the pixel i, I′ denotes a constrained region of an input image I, and |I′| denotes a number of pixels in I′.

5. The method according to claim 1 , wherein a segmentation loss is computed, for each of the first, second, and third segmentation losses, by:
seg = bce ( p,g )+ iou ( p,g )

wherein p denotes a probability map, g denotes a pseudo mask, bce denotes a binary cross entropy loss, and iou denotes an IoU loss.

6. The method according to claim 5 , wherein bce and iou are defined as:

Assignees

• Ping An Tech Shenzhen Co Ltd

Inventors

• Tang Youbao
• Cai Jinzheng
• Yan Ke
• Lu Le

Classifications

• G06T7/0012 Primary

  Biomedical image inspection · CPC title

• G06F18/213

  Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods · CPC title

• G06F18/253

  of extracted features · CPC title

• G06T3/40

  Scaling of whole images or parts thereof, e.g. expanding or contracting · CPC title

• G06T7/11

  Region-based segmentation · CPC title