Medical image analyzing system for identifying pancreatic cancer using machine learning and method thereof

Patent metadata

Field	Value
Publication number	US-12106473-B2
Application number	US-202117507949-A
Country	US
Kind code	B2
Filing date	Oct 22, 2021
Priority date	Oct 23, 2020
Publication date	Oct 1, 2024
Grant date	Oct 1, 2024

Abstract

A medical image analyzing system and a medical image analyzing method are provided and include inputting at least one patient image into a first model of a neural network module to obtain a result having determined positions and ranges of an organ and a tumor of the patient image; inputting the result into a second model of a first analysis module and a third model of a second analysis module, respectively, to obtain at least one first prediction value and at least one second prediction value corresponding to the patient image; and outputting a determined result based on the first prediction value and the second prediction value. Further, processes between the first model, the second model and the third model can be automated, thereby improving identification rate of pancreatic cancer.

First claim

What is claimed is:

1. A medical image analyzing system, comprising:
a neural network module having a first model and configured to input at least one patient image into the first model to obtain a result of determined positions and ranges of an organ and a tumor of the patient image;
a first analysis module having a second model and configured to input the result of the determined positions and ranges of the organ and the tumor of the patient image into the second model to obtain at least one first prediction value corresponding to the patient image;
a second analysis module having a third model and configured to input the result of the determined positions and ranges of the organ and the tumor of the patient image into the third model to obtain at least one second prediction value corresponding to the patient image;
a database stored with a plurality of images, organ position and range markers and tumor position and range markers, wherein the plurality of images, the organ position and range markers and the tumor position and range markers are interlinked to use as a first training set; and
a determining module configured to output a determined result based on the first prediction value and the second prediction value,
wherein the neural network module is trained to obtain the first model based on the first training set, wherein the neural network module is a model searched by using a coarse-to-fine neural Architecture search (C2FNAS), and wherein the neural network module uses an Adam optimizer and a cosine annealing learning rate scheduler to adjust a learning rate in a range of 10 −3 to 10 −5 , and a loss function is set to a Dice loss combined with a categorical cross-entropy loss.

2. The medical image analyzing system of claim 1 , wherein the neural network module obtains a result having determined positions and ranges of the organ and the tumor of the plurality of images by inputting the plurality of images into the first model and uses the result as a second training set.

3. The medical image analyzing system of claim 2 , wherein the first analysis module first performs 3D feature analysis on the second training set using an algorithm of radiomics to obtain a plurality of 3D feature values, and then trains a machine learning algorithm of a gradient boosting decision tree using the plurality of 3D feature values to obtain the second model.

4. The medical image analyzing system of claim 3 , wherein the first analysis module uses a plurality of filters to extract 3D features, and wherein the plurality of filters include 3 Laplacian of Gaussian filters, 8 wavelet transformation filters and 1 gradient filter.

5. The medical image analyzing system of claim 3 , wherein features selected by the radiomics include first order features, gray level co-occurrence matrix features, gray level dependence matrix features, gray level run length matrix features, gray level size zone matrix features, or neighboring gray tone difference matrix features.

6. The medical image analyzing system of claim 2 , wherein the second analysis module first performs 2D feature analysis on the second training set using the algorithm of radiomics to obtain a plurality of 2D feature values, and then trains a machine learning algorithm of a gradient boosting decision tree using the plurality of 2D feature values to obtain the third model.

7. The medical image analyzing system of claim 6 , wherein the second analysis module uses a plurality of filters to extract 2D features, and wherein the plurality of filters include 4 wavelet transformation filters and 1 gradient filter.

8. The medical image analyzing system of claim 6 , wherein features selected by the radiomics include first order features, gray level co-occurrence matrix features, gray level dependence matrix features, gray level run length matrix features, gray level size zone matrix features, or neighboring gray tone difference matrix features.

9. The medical image analyzing system of claim 1 , further comprising a threshold-value selection module configured to plot a curve for the first prediction value or the second prediction value, wherein a threshold value for determining whether there is cancer is determined from the curve, such that the first analysis module or the second analysis module determines whether the first prediction value or the second prediction value represents having cancer based on the threshold value.

10. The medical image analyzing system of claim 9 , wherein the curve is a receiver operating characteristic curve, and wherein the threshold value is a corresponding threshold value corresponding to a maximum value of a Youden index.

11. The medical image analyzing system of claim 10 , wherein the Youden index is calculated from a sensitivity and a specificity corresponding to each point in the curve according to a formula Youden index=sensitivity+specificity−1.

12. The medical image analyzing system of claim 1 , wherein the determining module uses an outcome probability of a logistic regression model as the determined result, and wherein the logistic regression model is obtained based on the first prediction value and the second prediction value.

13. The medical image analyzing system of claim 1 , wherein the determining module uses one or both of the first prediction value and the second prediction value representing having cancer as the determined result.

14. The medical image analyzing system of claim 1 , further comprising an image preprocessing module configured to process the patient image by resampling, windowing and normalization before inputting into the first model, the second model, or the third model.

15. A medical image analyzing method, comprising:
obtaining at least one patient image, and interlinking a plurality of images, organ position and range markers and tumor position and range markers to use as a first training set via a database stored with the plurality of images, the organ position and range markers and the tumor position and range markers;
inputting the patient image into a first model of a neural network module to obtain a result having determined positions and ranges of an organ and a tumor of the patient image, wherein the neural network module is trained to obtain the first model based on the first training set, wherein the neural network module is a model searched by using a coarse-to-fine neural Architecture search (C2FNAS), and wherein the neural network module uses an Adam optimizer and a cosine annealing learning rate scheduler to adjust a learning rate in a range of 10 −3 to 10 −5 , and a loss function is set to a Dice loss combined with a categorical cross-entropy loss;
inputting the result having determined positions and ranges of the organ and the tumor of the patient image into a second model of a first analysis module and a third model of a second analysis module, respectively, to obtain at least one first prediction value corresponding to the patient image and at least one second prediction value corresponding to the patient image; and
outputting a determined result by a determining module according to the first prediction value and the second prediction value.

16. The medical image analyzing method of claim 15 , wherein the neural network module obtains a result having determined positions and ranges of the organ and the tumor of the plurality of images by inputting the plurality of images into the first model and uses the result as a second training set.

17. The medical image analyzing method of claim 16 , wherein the first analysis module first performs 3D feature analysis on the second t

Assignees

• Univ Nat Taiwan

Inventors

• Wang Wei-Chung
• Liao Wei-Chih
• Liu Kao-Lang
• Chen Po-Ting
• Wang Po-Chuan
• Chang da-wei

Classifications

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06T2207/30204

  Marker · CPC title

• G06T2207/20064

  Wavelet transform [DWT] · CPC title

• G06T2207/20016

  Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform · CPC title