CN-122023868-A - Acute ischemic stroke functional outcome prediction method and device based on deep learning

CN122023868ACN 122023868 ACN122023868 ACN 122023868ACN-122023868-A

Abstract

The application relates to the technical field of deep learning, and discloses a method and a device for predicting the functional outcome of acute ischemic stroke based on deep learning; the method comprises the steps of utilizing a deep learning image segmentation model to segment a preprocessed stroke lesion image sequence and a preprocessed deep medullary vein image sequence, generating a first image sequence according to the preprocessed stroke lesion image sequence and a lesion segmentation mask sequence, generating a second image sequence according to the preprocessed deep medullary vein image sequence and the preprocessed deep medullary vein segmentation mask sequence, utilizing a deep learning classification model to respectively process the first image sequence and the second image sequence to obtain a first function ending prediction result and a second function ending prediction result, and generating a stroke function ending prediction result based on the first function ending prediction result and the second function ending prediction result. The application can more rapidly and accurately predict the functional ending of the stroke.

Inventors

QI SHOULIANG
ZANG PEIZHUO
JU RONGHUI
CHEN SHANNAN
ZHAO HUIHE
LIU LINGKAI
LI HONGYI
ZHOU BO

Assignees

东北大学
辽宁省人民医院

Dates

Publication Date: 20260512
Application Date: 20251218

Claims (10)

1. A method for predicting the functional outcome of acute ischemic stroke based on deep learning, comprising: acquiring a stroke lesion image sequence and a deep medullary vein image sequence of a target patient, and respectively preprocessing to obtain a preprocessed stroke lesion image sequence and a preprocessed deep medullary vein image sequence; Dividing the preprocessed stroke lesion image sequence by using a pre-trained deep learning image segmentation model to obtain a lesion segmentation mask sequence; dividing the preprocessed deep medullary vein image sequence to obtain a deep medullary vein division mask sequence; Multiplying the preprocessed stroke lesion image sequence with corresponding images in the lesion segmentation mask sequence to generate a first image sequence, and multiplying the preprocessed deep medullary vein image sequence with corresponding images in the deep medullary vein segmentation mask sequence to generate a second image sequence; Processing the first image sequence by using a first branch network in the pre-trained deep learning classification model to obtain a first function ending prediction result, and processing the second image sequence by using a second branch network in the pre-trained deep learning classification model to obtain a second function ending prediction result; generating an acute ischemic stroke functional outcome prediction result for the target patient based on the first functional outcome prediction result and the second functional outcome prediction result.
2. The method of claim 1, wherein the pre-trained deep-learning image segmentation model is generated according to the following method: Acquiring a first data set and a second data set, wherein the characteristics of each sample in the first data set are a sample stroke lesion image sequence, the labels of the samples are standard lesion mask sequences, the characteristics of each sample in the second data set are a sample deep medullary vein image sequence, and the labels of the samples are standard deep medullary vein mask sequences; And training an initial deep learning image segmentation model by using the first data set and the second data set to obtain the pre-trained deep learning image segmentation model.
3. The method according to claim 2, wherein the method further comprises: Taking each sample in the first data set as a first target sample, respectively performing first morphological operations on a sample stroke lesion image sequence and a corresponding standard lesion mask sequence of the first target sample to obtain a first enhanced stroke lesion image sequence set and a corresponding first enhanced standard lesion mask sequence set, wherein the first morphological operations comprise at least one of random horizontal overturning processing, erosion processing and expanding processing, and Respectively taking each sample in the second data set as a second target sample, and respectively carrying out second morphological operations on a sample deep medullary vein image sequence of the second target sample and a corresponding standard deep medullary vein mask sequence to obtain a first enhanced deep medullary vein image sequence set and a first enhanced standard deep medullary vein mask sequence set, wherein the second morphological operations comprise at least one of nonlinear deformation processing, amplification processing, shrinkage processing, erosion processing and turnover processing; The method comprises the steps of obtaining a trained data enhancement model, wherein the trained data enhancement model is obtained by training a generation countermeasure network model based on a ResViT model generator; Generating a second enhanced stroke lesion image sequence set and a corresponding second enhanced standard lesion mask sequence set based on the first enhanced stroke lesion image sequence set, the first enhanced standard lesion mask sequence set by using the trained data enhancement model, and Generating a second enhanced deep medullary vein image sequence set and a second enhanced standard deep medullary vein mask sequence set based on the first enhanced deep medullary vein image sequence set and the first enhanced standard deep medullary vein mask sequence set by using the trained data enhancement model; generating a new first data set according to the first data set, the second enhanced stroke lesion image sequence set and a second enhanced standard lesion mask sequence set; and generating a new second data set according to the second data set, the second enhanced deep medullary vein image sequence set and a second enhanced standard deep medullary vein mask sequence set.
4. The method of claim 2, wherein the initial deep-learning image segmentation model is implemented based on a denoising diffusion probability model, the initial deep-learning image segmentation model comprising a forward denoising process and a backward denoising process, wherein the backward denoising process is implemented based on a denoising module, the denoising module comprising a denoising-UNet network and a conditional-UNet network, wherein multi-level output characteristics of a decoder in the conditional-UNet network are used as input data of an encoder in the denoising-UNet network; training an initial deep learning image segmentation model by using the first data set and the second data set to obtain the pre-trained deep learning image segmentation model, wherein the training comprises the following steps: Respectively taking each sample in the first data set and the second data set as a sample to be processed, taking sample characteristics of the sample to be processed as input of the condition-UNet network, taking a sample tag as input of the forward noise adding process, and obtaining noise prediction data through processing of the forward noise adding process and the reverse noise removing process; and training the initial deep learning image segmentation model based on the noise prediction data by utilizing a preset objective function to obtain a trained model, and taking a denoising module in the trained model as the pre-trained deep learning image segmentation model.
5. The method of claim 4, wherein the encoder in the denoising-UNet network comprises a plurality of U-Net blocks, one Mamba block connected after each U-Net block.
6. The method of claim 2, wherein the pre-trained deep-learning classification model is generated according to the method of: The method comprises the steps of acquiring a third data set, wherein sample characteristics of samples in the third data set comprise a first sample image sequence and a second sample image sequence, and sample labels are real function ending data, the first sample image sequence is generated according to a lesion mask prediction sequence obtained by processing the sample stroke lesion image sequence by using the pre-trained deep learning image segmentation model, and the second sample image sequence is generated according to a deep medullary vein image sequence obtained by processing the sample deep medullary vein image sequence by using the pre-trained deep learning image segmentation model; And training an initial image classification model by using the third data set to obtain the pre-trained deep learning image classification model, wherein the initial image classification model comprises a first branch network and a second branch network which are respectively used for processing the first sample image sequence and the second sample image sequence.
7. The method of claim 6, wherein the sample characteristics of the samples in the third dataset further comprise sample demographic data and sample clinical data, wherein the sample demographic data comprises at least one of age, gender, the sample clinical data comprises at least one of clinical physiological indicators and clinical assessment scores; the method further comprises the steps of: Acquiring demographic data and clinical data of the target patient; processing the demographic data and the clinical data by utilizing a pre-constructed random forest model to obtain a third functional outcome prediction result; the generating an acute ischemic stroke functional outcome prediction result for the target patient based on the first functional outcome prediction result and the second functional outcome prediction result comprises: And generating the acute ischemic stroke function outcome prediction result according to the first function outcome prediction result, the second function outcome prediction result and the third function outcome prediction result.
8. The method of claim 7, wherein the clinical physiological index comprises at least one of systolic blood pressure, diastolic blood pressure, fasting blood glucose level, cholesterol level, total homocysteine level, and the clinical assessment score comprises at least one of NIHSS score, DMV score, DWI score.
9. The method of claim 7, wherein generating the acute ischemic stroke functional outcome prediction result based on the first functional outcome prediction result, the second functional outcome prediction result, and the third functional outcome prediction result comprises: determining a first basic probability distribution according to the first function solution prediction result; Determining a second basic probability distribution according to the second function solution prediction result; Determining a third basic probability distribution according to the third functional solution prediction result; and processing the first basic probability distribution, the second basic probability distribution and the third basic probability distribution based on a Dempster combination rule to obtain the acute ischemic stroke functional outcome prediction result.
10. An acute ischemic stroke functional outcome prediction device based on deep learning, characterized in that the device comprises: The acquisition module is used for acquiring a stroke lesion image sequence and a deep medullary vein image sequence of a target patient, and respectively preprocessing the stroke lesion image sequence and the deep medullary vein image sequence to obtain a preprocessed stroke lesion image sequence and a preprocessed deep medullary vein image sequence; The segmentation module is used for segmenting the preprocessed stroke lesion image sequence by utilizing a pre-trained deep learning image segmentation model to obtain a lesion segmentation mask sequence; The processing module is used for multiplying the preprocessed stroke lesion image sequence and the corresponding images in the lesion segmentation mask sequence to generate a first image sequence, and multiplying the preprocessed deep medullary vein image sequence and the corresponding images in the deep medullary vein segmentation mask sequence to generate a second image sequence; The prediction module is used for processing the first image sequence by utilizing a first branch network in the pre-trained deep learning classification model to obtain a first function ending prediction result, and processing the second image sequence by utilizing a second branch network in the pre-trained deep learning classification model to obtain a second function ending prediction result; And the fusion module is used for generating an acute ischemic stroke function outcome prediction result aiming at the target patient based on the first function outcome prediction result and the second function outcome prediction result.

Description

Acute ischemic stroke functional outcome prediction method and device based on deep learning Technical Field The application relates to the technical field of deep learning, in particular to a method and a device for predicting an acute ischemic stroke functional outcome based on deep learning. Background Acute Ischemic Stroke (AIS), a very damaging neurological disease affecting millions of people worldwide, is a major cause of morbidity, mortality, and long-term disability. The treatment methods of the strokes such as venous thrombolysis, intravascular thrombectomy and the like can reduce the death rate and improve the functional outcome after the strokes. However, AIS patients may suffer serious disability and even death even if diagnosed and treated in a timely manner. Thus, timely and accurate prediction of functional outcomes is critical to optimizing management strategies and improving patient outcomes. Currently, methods for predicting stroke functional outcomes are mainly performed by clinicians manually annotating SWI images and evaluating clinical signs and symptoms of patients using clinical assessment scales, such as the modified rank scale (mRS) and the National Institutes of Health Stroke Scale (NIHSS), to achieve prognosis prediction of AIS. However, this approach has certain limitations, 1, the approach is subjective, is prone to variability between observers, leading to inconsistent and potentially inaccurate assessment, and 2, the manual assessment is very time consuming, delaying the critical decision to manage the acute phase of stroke. Disclosure of Invention In view of the above, the embodiments of the present application provide a method and an apparatus for predicting an acute ischemic stroke functional outcome based on deep learning, which aim to solve the above-mentioned problems or at least partially solve the above-mentioned problems. In a first aspect, an embodiment of the present application provides a deep learning-based method for predicting an acute ischemic stroke functional outcome, the method comprising: acquiring a stroke lesion image sequence and a deep medullary vein image sequence of a target patient, and respectively preprocessing to obtain a preprocessed stroke lesion image sequence and a preprocessed deep medullary vein image sequence; Dividing the preprocessed stroke lesion image sequence by using a pre-trained deep learning image segmentation model to obtain a lesion segmentation mask sequence; dividing the preprocessed deep medullary vein image sequence to obtain a deep medullary vein division mask sequence; Multiplying the preprocessed stroke lesion image sequence with corresponding images in the lesion segmentation mask sequence to generate a first image sequence, and multiplying the preprocessed deep medullary vein image sequence with corresponding images in the deep medullary vein segmentation mask sequence to generate a second image sequence; Processing the first image sequence by using a first branch network in the pre-trained deep learning classification model to obtain a first function ending prediction result, and processing the second image sequence by using a second branch network in the pre-trained deep learning classification model to obtain a second function ending prediction result; generating an acute ischemic stroke functional outcome prediction result for the target patient based on the first functional outcome prediction result and the second functional outcome prediction result. In a second aspect, an embodiment of the present application further provides an acute ischemic stroke functional outcome prediction device based on deep learning, where the device includes: The acquisition module is used for acquiring a stroke lesion image sequence and a deep medullary vein image sequence of a target patient, and respectively preprocessing the stroke lesion image sequence and the deep medullary vein image sequence to obtain a preprocessed stroke lesion image sequence and a preprocessed deep medullary vein image sequence; The segmentation module is used for segmenting the preprocessed stroke lesion image sequence by utilizing a pre-trained deep learning image segmentation model to obtain a lesion segmentation mask sequence; The processing module is used for multiplying the preprocessed stroke lesion image sequence and the corresponding images in the lesion segmentation mask sequence to generate a first image sequence, and multiplying the preprocessed deep medullary vein image sequence and the corresponding images in the deep medullary vein segmentation mask sequence to generate a second image sequence; The prediction module is used for processing the first image sequence by utilizing a first branch network in the pre-trained deep learning classification model to obtain a first function ending prediction result, and processing the second image sequence by utilizing a second branch network in the pre-trained deep learning classification model to obtain