Pancreatic postoperative diabetes prediction system based on supervised deep subspace learning

Patent metadata

Field	Value
Publication number	US-12381007-B2
Application number	US-202418788009-A
Country	US
Kind code	B2
Filing date	Jul 29, 2024
Priority date	Apr 29, 2022
Publication date	Aug 5, 2025
Grant date	Aug 5, 2025

Abstract

A pancreatic postoperative diabetes prediction system based on supervised deep subspace learning. A deep convolutional neural network and the MITK software are used to obtain postoperative residual pancreas area, so as to taken as the region-of-interest. Traditional image radiomics features and deep semantic features are extracted from the residual pancreas area, and a high-dimensional image feature set is constructed. Clinical factors related to diabetes, including pancreatic excision rate, fat and muscle tissue components, demographic information and living habits are extracted, and a clinical feature set is constructed. Based on a supervised deep subspace learning network, image and clinical features are represented and fused in subspace in dimensionality reduction, while a prediction model is trained to mine sensitive features highly relevant to the prediction risk of a patient suffering postoperative diabetes mellitus with a high degree of automation and discriminative accuracy.

First claim

What is claimed is:

1. A pancreatic postoperative diabetes prediction system based on supervised deep subspace learning, comprising a preoperative CT image data acquisition module, a residual pancreatic region of interest acquisition module, an image feature calculation module, a clinical feature calculation module and a deep subspace learning module;
wherein the preoperative CT image data acquisition module is configured to acquire CT image data before a pancreatectomy, and input the CT image data into the residual pancreatic region of interest acquisition module and the image feature calculation module;
wherein the residual pancreatic region of interest acquisition module is configured to input preoperative CT image data into a trained pancreas segmentation network to obtain a pancreas prediction region; in the pancreas prediction region, an excised edge of a pancreas is simulated by software to obtain a residual pancreatic region after the pancreatectomy, and input the residual pancreatic region, as a region of interest for a subsequent calculation of an image feature, into the image feature calculation module;
wherein the image feature calculation module is configured to calculate a pancreatic image feature according to the preoperative CT image data and a region of interest of the image feature, and input the pancreatic image feature to the deep subspace learning module;
wherein the clinical feature calculation module is configured to acquire clinical information related to postoperative diabetes of a patient, comprising demographic information, pancreatic volume resection rate, pancreatic residual volume and abdominal fat and muscle content features, perform feature concatenation to form a clinical feature, and input the clinical feature into the deep subspace learning module; and
wherein the deep subspace learning module is configured to perform feature dimensionality reduction and fusion through a deep subspace learning network, the deep subspace learning network comprises an encoder, a latent spatial variable self-representative layer and a decoder, and for supervising a learning of the latent spatial variable self-representative layer; the deep subspace learning network inputs the pancreatic image feature and the clinical feature, outputs a latent spatial variable through the encoder, connects the latent spatial variable output by the encoder with a fully connected layer, and effects an activation function to obtain a predicted value of diabetes risk; a loss function of the deep subspace learning network in the deep subspace learning module is:

L
⁡
(
Θ
)

=



X
-

X
^



F
2

+

2
⁢

Tr
⁡
(

X
T

⁢
L
⁢

X
^

)

+

1
2

⁢

γ
1

⁢



Z
-
ZC



F
2

+

α
2

⁢

γ
1

⁢


C


1

+

γ
2

⁢

BCE
⁡
(

y
^

,
y

)

,

s
.
t
.

diag
⁡
(
C
)

=
0

where X={X img , X clinic }, X img represents the image feature, X clinic represents the clinical feature, {circumflex over (X)} represents an output of the decoder, y represents a real situation of the patient suffering from postoperative diabetes, ŷ represents a diabetes risk predicted by a model, Z represents the latent spatial variable output by the encoder, L represents a Laplacian matrix, a symbol Tr represents a trace of a matrix, and a symbol T represents a matrix transposition, Θ represents all parameters in the deep subspace learning network, comprising an encoder parameter Θ e , a self-representative coefficient matrix C, a supervision module parameter Θ s and a decoder parameter Θ d ; α, γ 1 and γ 2 are regularization coefficients, a symbol ∥·∥F represents a frobenius norm, and BCE(·) represents a cross entropy loss.

2. The pancreatic postoperative diabetes prediction system based on supervised deep subspace learning according to claim 1 , wherein the preoperative CT image data acquisition module truncates a HU value of the CT image data to [−100, 240] and discretizes the HU value to [0,255] after acquiring the CT image data before the pancreatectomy, calculates a rectangular frame of a region surrounded by residual pancreas, sets an edge expansion value, and truncates a rectangular frame of the CT image data and a residual pancreas labeled image.

3. The pancreatic postoperative diabetes prediction system based on supervised deep subspace learning according to claim 1 , wherein in the residual pancreatic region of interest acquisition module, a preoperative pancreatic CT image is automatically segmented based on the deep convolutional neural network to obtain a complete pancreas prediction region, and a surgical cutting plane is simulated in Medical Imaging Interaction Toolkit (MITK) software according to a surgical rec

Assignees

• Zhejiang Lab

Inventors

• Li Jingsong
• Hu Peijun
• Tian Yu
• Zhou Tianshu

Classifications

• G06T2207/30092

  Stomach; Gastric · CPC title

• G06T2207/20221

  Image fusion; Image merging · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06T2207/10081

  Computed x-ray tomography [CT] · CPC title