TWI895233B - Deep learning method and identification device for identifying prostate cancer - Google Patents

Abstract

A deep learning method and recognition device for identifying prostate cancer, the deep learning method comprising the following steps: (a) obtaining a plurality of image data for generating an image, each of the image data including level information related to the degree of suspicion of a prostate cancer lesion. (b) extracting a feature signal set from each of the image data. (c) performing an artificial intelligence learning algorithm based on the level information, feature signal set, and an optimized parameter set for each of the image data to establish a recognition model. Thus, by using pre-classified image data, the processing speed is accelerated, and by using a limited numerical optimized parameter set, the recognition model's accuracy in identifying prostate cancer is significantly improved to over 90%.

Description

Deep learning method and identification device for identifying prostate cancer

The present invention relates to an identification device, and more particularly to a deep learning method and an identification device for identifying prostate cancer.

A known deep learning system for identifying prostate cancer bone metastases based on whole-body bone scans, disclosed in Republic of China Patent Publication No. TW202117742A, includes a pre-processing module that receives and processes whole-body bone scans and a neural network module that detects whether the whole-body bone scans represent prostate cancer bone metastases. The neural network module uses a Faster R-CNN (Fast R-CNN) model to segment the thoracic region of the training image based on the whole-body bone scan input and classify cancer cell bone metastases.

However, the model exhibits a tendency to overfit the training data and lose its generalization capabilities. In practical applications, it faces technical issues such as being unable to adapt to new data and its accuracy being lower than expected.

Furthermore, the whole-body bone scan image requires segmentation of the chest area into a training image, which is not only costly but also time-consuming.

In addition, training large models requires a large amount of GPU/TPU resources, which creates a barrier for small and medium-sized enterprises or developers.

Therefore, the purpose of the present invention is to provide a deep learning method and recognition device for identifying prostate cancer that can accelerate processing speed and effectively improve recognition accuracy.

Therefore, the deep learning method for identifying prostate cancer of the present invention includes the following steps:

(a) Obtaining a plurality of image data for generating an image, each of the image data including level information related to the suspicion level of a prostate cancer lesion.

(b) extracting a feature signal group from each of the image data.

(c) performing an artificial intelligence learning algorithm based on the level information of each image data, the feature signal set, and an optimization parameter set to establish a recognition model, the optimization parameter set including an initial learning rate, a learning rate drop period, a learning rate drop factor, and an L2 regularization, wherein the initial learning value is configured as 0.001, the learning rate drop period is configured as 10, the learning rate drop factor is configured as 0.1, and the L2 regularization is configured as 0.0001.

(d) outputting risk assessment information based on the recognition model and each feature signal set of the image data.

An identification device that implements the aforementioned deep learning method for identifying prostate cancer includes a communication module and a processing module.

The communication module is used to obtain the image data and output the recognition model.

The processing module is connected to the communication module and is used to extract the characteristic signal set of each image data, and perform artificial intelligence learning calculation based on the level information of each image data, the characteristic signal set, and the optimization parameter set to establish the recognition model.

A prostate cancer identification device includes a communication module and a processing module.

The communication module is used to load the recognition model mentioned above, obtain an image to be recognized, and output another risk assessment information.

The processing module is connected to the communication module, and uses the recognition model to recognize the image to be recognized and generate the other risk assessment information.

The invention's effectiveness lies in accelerating processing speed by using pre-classified image data and significantly increasing the accuracy of the recognition model in identifying prostate cancer to over 90% by using an optimized parameter set with limited values.

Referring to FIG. 1 , an embodiment of the identification device for identifying prostate cancer according to the present invention includes a communication module 1 , a processing module 2 , and a display module 3 .

The identification device typically uses the Internet as a medium and possesses powerful computing capabilities and ample storage space. The identification device is not limited to a specific number of machines, but rather refers to equipment capable of completing a large amount of work in a short period of time and providing services to a large number of users. In this embodiment, it can be a server, laptop, desktop computer, or other device capable of inputting or outputting data and information.

The communication module 1 is capable of communicating with other devices via wireless, mobile, or wired communication technologies. In this embodiment, the wireless communication technology is configured as Wi-Fi, Bluetooth, NFC, or a combination thereof. The mobile communication technology is configured as 4G, 5G, or a combination thereof. The wired communication technology is configured as USB, Type-C, or a combination thereof.

The processing module 2 is connected to the communication module 1 and is responsible for executing various instructions and data processing tasks, such as graphics rendering, image processing, deep learning training, etc. In this embodiment, the processing module 2 can be a CPU module, a GPU module, a TPU module, or a combination thereof.

The display module 3 is connected to the processing module 2 and is used to display text, images, videos, or a combination thereof.

Referring to Figures 1, 2, and 3, the deep learning method for identifying prostate cancer of the present invention performs the following steps through the processing module 2 of the identification device:

Step S01: Obtain a plurality of image data D for generating an image through the communication module 1. Each image data D includes level information related to the suspicion level of a prostate cancer lesion.

In this embodiment, the image data D acquired in step S01 is used to generate MRI images (magnetic-resonance imaging), CT images (computed tomography), US images (Ultrasound), or a combination thereof. The image data D may be sourced from the PROSTATEx database, the Prostate-MRI-US-Biopsy database, or other publicly available medical imaging databases.

The grade information of each image data D includes a grade number. The grade number ranges from 1 to n. A larger grade number indicates a higher risk. In this embodiment, the grade information of each image data D is classified into grades 1 to 5 according to the Gleason grading system based on clinical diagnostic standards.

Step S02: Extract a feature signal group from each image data D.

In this embodiment, a convolutional neural network (CNN) based on the MobileNet-V2 architecture is used to automatically extract the feature signal set.

Step S03: Based on the level information, feature signal set, and optimization parameter set of each image data D, a convolutional neural network with a MobileNet-V2 architecture and an optimizer are used to perform artificial intelligence learning calculations to establish a recognition model M.

In this embodiment, the optimizer is configured as an ADAM optimizer, an RMSprop optimizer, or an SGDM optimizer.

The optimization parameter set includes an initial learning rate, a learning rate drop period, a learning rate drop factor, an L2 regularization, and a squared gradient decay factor.

The initial learning value is configured as 0.001 to control the weight update amplitude of the recognition model M in the early stage of training. If the learning rate is too large, the recognition model M may converge unstably, while if it is too small, it may converge too slowly.

This learning rate reduction cycle can be combined with a staged learning rate to periodically reduce the learning rate, allowing the recognition model M to more steadily approach the optimal recognition rate. In this embodiment, the learning rate reduction cycle is configured to 10, and the learning rate reduction coefficient is configured to 0.1. This allows the learning rate of the recognition model M to gradually decrease over time from the initial learning value. For example: in the first cycle, the learning rate = 0.001 × 0.1 = 0.0001; in the second cycle, the learning rate = 0.0001 × 0.1 = 0.00001.

The L2 regularization value is configured to 0.0001, which can limit the growth of weights, prevent overfitting, and improve the generalization ability of the model.

The squared gradient attenuation coefficient is configured to be less than 1, preferably 0.9. By maintaining the sum of squared gradients, the learning rate is adjusted to achieve an adaptive learning rate. This prevents the sum of squared gradients from being too large, thus preventing the learning rate from being too small, causing the training process to stagnate.

Step S04: Output risk assessment information S based on the recognition model M and the characteristic signal set of each image data D.

In this embodiment, the risk assessment information S includes a risk level. The risk level is configured as 5 levels and is presented in text as very low, low, medium, high, and very high.

This risk level is very low, meaning no obvious abnormalities or lesions are observed and there is no risk of cancer.

This risk rating is low, meaning some abnormalities were observed but are benign.

This risk level is moderate, meaning some abnormalities were observed and the likelihood of cancer is low.

The risk level is high, meaning some abnormalities were observed and the possibility of cancer is higher.

This risk level is extremely high, meaning a significant abnormality has been observed and the possibility of cancer is very high.

Table 1: Architecture Optimizer VGG-16 MobileNet-V2 Inception-V3 ADAM 93.6% 93.2% 93.6% RMSprop 93.67% 95.07% 70.73% SGDM 78.93% 94.13% 93.2%

When training and testing the recognition model M, the processing module 2 divides the image data D into a training set DS1, a test set DS2, and a validation set DS3. The image data D in the training set DS1 accounts for 80%. The image data D in the test set DS2 accounts for 10%, and the image data D in the validation set DS3 accounts for 10%. As shown in Table 1, the convolutional neural network (CNN) with the MobileNet-V2 architecture can achieve an accuracy of over 90% when combined with different optimizers. Among them, the optimizer with the RMSprop architecture can achieve an accuracy of up to 95.07%.

Thus, another recognition device installed in a medical institution only needs to load the recognition model M through the communication module 1 and obtain an image P to be recognized. The processing module 2 of the other recognition device can then use the recognition model M to recognize the image P to be recognized, generate risk assessment information S, and display this risk assessment information S through the display module 3. This helps doctors diagnose prostate cancer lesions and provides more accurate tumor monitoring and treatment recommendations.

It is worth noting that the image to be recognized P can also serve as a training image. This increases the amount of image data D in the training set DS1, allowing for continuous improvement of the algorithm and the recognition model M to enhance learning effectiveness and recognition rate.

It should be noted that the identification device installed in the medical institution and the trial identification device for establishing the identification model M are not limited to being two different devices. In other variations of this embodiment, they can also be the same device.

Based on the above description, the advantages of the aforementioned embodiments can be summarized as follows:

1. By using a limited set of optimized parameters, the present invention significantly improves the accuracy of prostate cancer recognition model M to over 90%. When a convolutional neural network based on the MobileNet-V2 architecture is used in conjunction with an optimizer based on the RMSprop architecture, an accuracy of 95.07% is achieved. Clearly, this invention significantly improves recognition accuracy compared to learning techniques.

2. It is also discovered that using pre-classified image data D can not only speed up processing but also improve the learning effect when establishing the recognition model M.

3. The convolutional neural network of the MobileNet-V2 architecture is a lightweight model suitable for resource-limited environments.

However, the above description is merely an example of the present invention and should not be used to limit the scope of the present invention. All simple equivalent changes and modifications made within the scope of the patent application and the contents of the patent specification of the present invention are still within the scope of the present patent.

1: Communication module 2: Processing module 3: Display module D: Image data DS1: Training Set DS2: Test set DS3: Verification set M: Recognition model P: Image to be recognized S: Risk assessment information

Other features and functions of the present invention will be clearly presented in the embodiments with reference to the accompanying drawings, wherein: FIG1 is a block diagram illustrating an embodiment of an identification device for identifying prostate cancer according to the present invention; FIG2 is a schematic diagram illustrating a display module for displaying risk assessment information in the embodiment; and FIG3 is a flow chart of the deep learning method for identifying prostate cancer according to the present invention.

1: Communication module

2: Processing module

3: Display module

D: Image data

DS1: Training set

DS2: Test set

DS3: Verification Set

M: Identification Model

P: Image to be recognized

S: Risk Assessment Information

Claims (10)

A deep learning method for identifying prostate cancer comprises the following steps: (a) obtaining a plurality of image data for generating an image, each of the image data comprising level information related to the degree of suspicion of a prostate cancer lesion; (b) extracting a feature signal set from each of the image data; and (c) performing an artificial intelligence learning algorithm based on the level information of each of the image data, the feature signal set, and an optimization parameter set to establish a recognition model, wherein the optimization parameter set comprises an initial learning rate, a learning rate drop period, a learning rate drop factor, and an L2 regularization (L2 Regularization), wherein the initial learning value is configured as 0.001, the learning rate reduction cycle is configured as 10, the learning rate reduction coefficient is configured as 0.1, and the L2 regularization is configured as 0.0001; and (d) outputting risk assessment information based on the recognition model and the feature signal group of each of the image data. The deep learning method for identifying prostate cancer as described in claim 1, wherein the image data obtained in step (a) is used to generate MRI images (magnetic-resonance imaging), CT images (computed tomography), US images (Ultrasound), or a combination thereof. The deep learning method for identifying prostate cancer as described in claim 1, wherein the level information of the image data in step (a) includes a level number, the level number is between 1 and n, and a larger level number indicates a higher risk. A deep learning method for identifying prostate cancer as described in claim 1, wherein step (b) and step (c) use a convolutional neural network (CNN) with a MobileNet-V2 architecture to extract a feature signal set for each of the image data and establish the recognition model. The deep learning method for identifying prostate cancer as described in claim 4, wherein step (c) further uses an optimizer to establish the recognition model, and the optimizer can be an ADAM optimizer, an RMSprop optimizer, or an SGDM optimizer. The deep learning method for identifying prostate cancer as described in claim 5, wherein the optimization parameter set further includes a squared gradient decay factor (SGF), and the squared gradient decay factor is configured to be less than 1. The deep learning method for identifying prostate cancer as described in claim 6, wherein step (c) uses an RMSprop optimizer to establish the recognition model, and the squared gradient attenuation coefficient is configured to be 0.9. An identification device that executes the deep learning method for identifying prostate cancer as described in any one of claims 1 to 7, comprising: a communication module for obtaining the image data and outputting the recognition model; and a processing module connected to the communication module, for extracting a feature signal set of each image data, performing an artificial intelligence learning algorithm based on the level information of each image data, the feature signal set, and the optimization parameter set to establish the recognition model. A recognition device for identifying prostate cancer comprises: a communication module for loading a recognition model established by the deep learning method for identifying prostate cancer as described in claim 1, obtaining an image to be recognized, and outputting another risk assessment information; and a processing module connected to the communication module, using the recognition model to recognize the image to be recognized and generate the another risk assessment information. The identification device for identifying prostate cancer as described in claim 9 further includes a display module. The communication module loads the identification model, obtains the image to be identified, and outputs the other risk assessment information through wireless communication technology or wired communication technology. The display module is connected to the processing module and is used to display the other risk assessment information.