A method for classifying clinical pregnancy and biochemical pregnancy outcome based on static images

Abstract

This invention discloses a method for predicting the outcome of in vitro fertilization-embryo transfer (IVF-ET) based on static embryo images, used to distinguish between biochemical pregnancy and clinical pregnancy. It belongs to the field of medical artificial intelligence and assisted reproductive technology and is applicable to single embryo transfer (SET) and dual embryo transfer (DET) scenarios. The method includes: acquiring a static image of at least one embryo to be transferred; preprocessing and standardizing the embryo image, and then inputting it into a feature extraction model based on a convolutional neural network. The model uses DenseNet-169 as the backbone network and combines it with a Feature Pyramid Network (FPN) for multi-scale feature fusion. When the input is a dual embryo image, a Trophectoderm–Inner Cell Attention (TEIC-Attn) attention module is introduced to adaptively weight and fuse the features of the two embryos to characterize the contribution of different embryos to the final transfer outcome. When the input is a single embryo image, the prediction is completed based on the corresponding embryo features. The model ultimately outputs a probability prediction result of whether the IVF treatment outcome is a biochemical pregnancy or a clinical pregnancy. Furthermore, this invention combines a gradient-based class activation mapping method to generate interpretable visual heatmaps, highlighting key morphological regions closely related to embryo transfer outcomes, such as the inner cell mass and trophectoderm, thereby providing clinicians and patients with intuitive and reliable decision-making support. This invention enables unified prediction and interpretation of single and double embryo transfer outcomes without relying on time-series data, improving the accuracy and interpretability of in-vitro fertilization treatment decisions and demonstrating promising clinical application prospects.

Description

Technical Field

[0001] This invention belongs to the field of medical artificial intelligence and assisted reproductive technology, specifically relating to a deep learning-based method for classifying early pregnancy outcomes in in vitro fertilization-embryo transfer (IVF-ET). This method uses static embryo images to distinguish between biochemical and clinical pregnancies and is applicable to both single embryo transfer (SET) and double embryo transfer (DET) scenarios. Background Technology

[0002] Infertility has become an increasingly prominent public health problem worldwide. Statistics from the World Health Organization show that approximately one in six couples of reproductive age experience infertility to varying degrees. In vitro fertilization-embryo transfer (IVF-ET), as one of the most effective clinical methods for treating infertility, involves fertilizing gametes and culturing early embryos outside the body, then transferring the embryos back into the mother's uterus to achieve pregnancy. It is widely used in clinical practice.

[0003] In IVF clinical practice, early pregnancy outcomes after embryo transfer mainly fall into two categories: clinical pregnancy and biochemical pregnancy. Clinical pregnancy refers to a persistent pregnancy confirmed by ultrasound examination of a gestational sac within the uterine cavity, typically indicating a high probability of continued pregnancy and live birth. Biochemical pregnancy, on the other hand, is an early pregnancy loss confirmed only by elevated serum human chorionic gonadotropin (β-hCG) levels, but without a visible gestational sac on ultrasound. Accurately distinguishing between these two outcomes is of significant clinical importance for objectively assessing the true developmental potential of the embryo, optimizing embryo transfer decisions, reducing the psychological burden on patients, and rationally allocating medical resources.

[0004] Currently, clinical practice primarily relies on embryo morphology assessment systems (such as the Gardner scoring system) combined with embryologists' experience for embryo selection and pregnancy prognosis assessment. However, this method still has significant limitations: firstly, morphological assessment is highly dependent on human experience, resulting in low consistency among different embryologists; secondly, traditional scoring systems struggle to quantify and capture subtle morphological characteristics of embryos and their potential biological differences; and thirdly, existing assessment systems have limited predictive ability for early pregnancy outcomes, especially biochemical pregnancies, making it difficult to reliably determine whether a pregnancy can continue before embryo transfer.

[0005] With the development of artificial intelligence technology, deep learning methods have made significant progress in the field of image processing, and have also provided new technical means for embryo evaluation and outcome prediction in the field of assisted reproduction. Models based on Convolutional Neural Networks (CNNs) can automatically learn multi-level features from raw images, demonstrating superior performance compared to traditional machine learning methods in medical image classification and analysis tasks. Densely connected convolutional networks, represented by DenseNet, have achieved good application results in various medical image analysis scenarios by enhancing feature reuse and gradient propagation capabilities. Meanwhile, Feature Pyramid Networks (FPNs) can improve the model's ability to recognize complex morphological structures by fusing feature information at different scales; the Attention Mechanism, through adaptive weighting of key features, helps enhance the model's ability to focus on important structural regions, and has high application value in medical image analysis tasks with limited sample size.

[0006] Currently, several artificial intelligence-based IVF outcome prediction methods have been proposed and authorized, such as: a multimodal embryo pregnancy outcome prediction device (CN109544512B), an embryo developmental potential prediction method, system, device, and storage medium (CN113469958B), a deep learning-based method for extracting embryo morphological and dynamic parameters and predicting blastocysts (CN117612160A), and an automated embryo morphology assessment system based on artificial intelligence image recognition (CN120108720A). However, the aforementioned existing technologies still generally suffer from the following shortcomings: (1) The definition of pregnancy outcome is crude. Most existing methods set the prediction target as "pregnancy / non-pregnancy" or directly use "live birth" as the endpoint, without distinguishing between clinical pregnancy and biochemical pregnancy. These two outcomes are fundamentally different in biological significance and clinical value. Embryos that can lead to biochemical pregnancy are often regarded as "successful pregnancy" in the above models, but this does not meet the actual clinical need for continued pregnancy. Predictive models that distinguish this precise outcome are still lacking.

[0007] (2) Difficult to adapt to double embryo transfer scenarios. Most existing models are designed only for single embryo transfer, while in real clinical practice, double embryo transfer is still widely used in specific patient populations or to improve the cumulative pregnancy rate. The pregnancy outcome of double embryo transfer is the result of the combined effect of the two embryos. Existing models cannot process double embryo inputs at the same time, nor can they assess the relative contribution of different embryos to the final outcome, which limits their application value in real clinical scenarios.

[0008] (3) Lack of interpretability mechanisms. Some existing deep learning-based embryo assessment methods do not incorporate interpretability analysis mechanisms, and the model prediction process is characterized by "black box" features. Clinicians cannot know the embryonic structural or morphological regions on which the model's decisions are based. This lack of transparency in prediction methods is inconsistent with the requirement for clear evidence in clinical decision-making, which seriously restricts the clinical trust and widespread application of artificial intelligence models in the field of assisted reproduction.

[0009] (4) High dependence on expensive imaging equipment. Some methods rely on time-lapse imaging systems to extract embryo morphological and dynamic parameters. Although this adds time dimension information, the related equipment is expensive and not yet widely used in most reproductive centers. In contrast, static embryo images are the most commonly collected, lowest-cost, and most universal data form in current IVF laboratories. Constructing accurate and interpretable prediction methods based on static images has greater clinical application value.

[0010] Therefore, there is an urgent need for an IVF early pregnancy outcome prediction method that can uniformly model single and double embryo transfer scenarios based on static embryo images, distinguish between clinical and biochemical pregnancies, and provide interpretability. Summary of the Invention

[0011] To overcome the shortcomings of existing technologies in predicting early pregnancy outcomes in in vitro fertilization-embryo transfer (IVF-ET), this invention aims to provide a method for distinguishing between clinical and biochemical pregnancy outcomes based on static embryo images. This method utilizes an interpretable deep learning model to predict the pregnancy outcome type of the embryo before transfer, and is applicable to both single and double embryo transfer scenarios. It provides clinicians with objective and interpretable criteria for embryo selection, thereby improving the scientific rigor and reliability of assisted reproductive technology decisions.

[0012] To achieve the above objectives, the present invention provides a method for distinguishing between clinical pregnancy and biochemical pregnancy outcomes based on static embryo images, comprising the following steps: S1: Acquire static image data of the embryo to be transferred, the image data including single embryo images or double embryo images; S2: Preprocess the acquired embryo images to obtain standardized input images; S3: Input the preprocessed image into the feature extraction module. The feature extraction module includes a convolutional neural network with DenseNet169 as the backbone, and combines it with a feature pyramid network to perform multi-scale feature extraction on the embryo image to obtain the corresponding depth feature vector. S4: Input the deep feature vector into the attention fusion module to adaptively fuse the features of single or dual embryos and generate the corresponding fused features; S5: Input the fused features into the classification module and output the probability prediction results of the embryo transfer outcome being clinical pregnancy or biochemical pregnancy; S6: Classify the probability prediction results according to the set judgment rules to distinguish between clinical pregnancy and biochemical pregnancy outcomes.

[0013] In the aforementioned technical solution, static embryo images acquired from real clinical settings are used as input, and a unified model architecture supports predictions for both single and double embryo transfer scenarios. The model uses DenseNet169 as its core feature extractor, combined with a feature pyramid network to enhance the capture of morphological features at different developmental stages, and introduces a specialized attention mechanism to achieve differentiated fusion and interpretability weighting of features from double embryos. The final model not only outputs pregnancy outcome predictions but also provides visual interpretation through gradient-weighted class activation mapping technology, highlighting key morphological regions of the embryo that the model's decisions focus on.

[0014] This predictive method directly differentiates between clinical and biochemical pregnancies early on, without relying on ultrasound results, and can provide outcome predictions before embryo transfer. The model exhibits good interpretability; visualization of attention weights helps understand the contribution of each embryo in a twin pregnancy, enhancing the transparency and reliability of clinical applications. After training, prediction results can be quickly obtained simply by inputting static images of the embryos, providing important reference for embryo selection and transfer decisions.

[0015] The core formula for the attention fusion module in step 4) is:

[0016]

[0017] in, and Given two feature vectors as input, This represents a vector concatenation operation. and These are the learnable weights and biases of the first and second fully connected layers, respectively. , For the generated attention weights, The final output is the fused feature vector.

[0018] The Feature Pyramid Network is used to address the multi-scale feature representation problem in embryo image analysis. During embryonic development, embryos at different stages exhibit significant differences in morphological structure, size, spatial resolution, and semantic level. The Feature Pyramid Network, through top-down feature transfer and lateral connection structures, fuses high-level semantic features with low-level detailed features, thereby constructing a multi-scale feature representation that combines semantic information and spatial resolution. This allows the model to simultaneously focus on the overall morphological features of the embryo and local cellular structures, improving the accuracy of pregnancy outcome prediction.

[0019] While predicting pregnancy outcomes, this invention also visualizes the model's decision-making process using a gradient-based class activation mapping method, generating corresponding interpretable heatmaps to highlight key embryonic morphological regions that the model focuses on during the prediction process, including structures such as the inner cell mass and trophectoderm.

[0020] In the case of twin embryo transfer, the output of the attention weight can intuitively reflect the relative contribution of different embryos to the prediction results, thereby providing clinicians with an understandable and traceable auxiliary reference in the process of embryo selection and transfer decision-making.

[0021] Compared with the prior art, the present invention has at least the following beneficial effects: (1) This invention is the first to achieve direct differentiation between clinical pregnancy and biochemical pregnancy outcomes based on static embryo images, making up for the shortcomings of the overly coarse prediction targets of pregnancy outcomes in the prior art; (2) This invention supports both single embryo transfer and double embryo transfer scenarios through a unified model architecture, and achieves adaptive fusion of double embryo features through an attention mechanism, which significantly improves the applicability of the method in real clinical application scenarios. (3) While outputting the pregnancy outcome prediction results, the present invention can provide embryo-level weights and corresponding interpretable visual information, which enhances the transparency and clinical credibility of the model prediction process. (4) This invention relies only on static embryo image data and does not require additional expensive imaging equipment, thus having good clinical promotion value and application prospects. Attached Figure Description

[0022] Figure 1 is a flowchart of the model in this embodiment of the invention when predicting IVF-ET treatment outcomes; Figure 2 shows the image feature extraction module in an embodiment of the present invention; Figure 3 shows the initial feature extraction module block0 of the dense convolution; Figure 4 shows the intermediate modules for dense convolutional feature extraction, blocks 1-3. Figure 5 shows the attention mechanism module after multi-scale feature fusion; Figure 6 is a structural diagram of the attention mechanism; Figure 7 shows the classification head module after the attention mechanism. Detailed Implementation Plan

[0023] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the specific embodiments of this invention will be described in detail below with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0024] like Figure 1 The diagram illustrates the overall process of a predictive model used in an embodiment of the present invention to distinguish between clinical and biochemical pregnancy outcomes. This embodiment provides a method for distinguishing pregnancy outcomes based on static embryo images, specifically including the following steps: Step 101: Obtain embryo images. Acquire static embryo image data from the patient. The embryo images can be a single embryo image in a single embryo transfer scenario or two embryo images in a double embryo transfer scenario. The system identifies the number of input embryo images and maps them to a unified feature processing interface to ensure the consistency of the subsequent model structure.

[0025] Step 102, Image Preprocessing. The acquired embryo images undergo standardized preprocessing. First, an automatic segmentation model is used to segment the embryo images foreground to remove background regions and retain the main embryo. Then, the segmented embryo images are normalized, and key regions are cropped. Finally, the processed images are uniformly scaled to a preset standard resolution to meet the input requirements of the feature extraction network. During the model training phase, data augmentation operations can be performed on the training samples to improve the model's generalization ability.

[0026] Step 103, Multi-scale Feature Extraction. Input the preprocessed embryo image into the image feature extraction module. For example... Figure 2 As shown, the feature extraction module uses the DenseNet169 network as its backbone structure and combines it with a feature pyramid network to perform multi-scale feature modeling of embryo images. Figure 3 As shown, the initial feature extraction module includes a convolutional layer, a normalization layer, an activation layer, and a pooling layer, used to initially encode the input image and reduce its spatial resolution. Figure 4 As shown, the DenseNet169 backbone network consists of multiple sequentially connected DenseBlocks and transition layers. Each DenseBlock is composed of several densely connected convolutional units, used to extract embryonic morphological features from low-level texture to high-level semantics layer by layer. The feature pyramid network fuses feature maps from different levels through a top-down path and lateral connection structure, thereby obtaining multi-scale feature representations with both high resolution and strong semantic expressive power.

[0027] Step 104, Attention Fusion Module (TEIC-Attn) Processing. The embryonic features extracted in Step 103 are input into the attention fusion module. For example... Figure 5 As shown, the attention fusion module is used to adaptively weight and fuse features from single or dual embryos, and outputs the fused features and corresponding attention weights. Figure 6 As shown, when two embryonic feature vectors exist, the attention fusion module first concatenates the feature vectors, and then generates an attention weight vector through a lightweight neural network to characterize the relative contribution of different embryonic features to the prediction result. Finally, the module performs a weighted summation of each embryonic feature according to the attention weights to obtain a unified fused feature vector.

[0028] Step 105, classification prediction. For example... Figure 7 As shown, the fused feature vector output by the attention fusion module is input into the classification module. The classification module includes at least one fully connected layer, a non-linear activation layer, and an output layer, used to map the fused features into predicted probability values ​​of pregnancy outcomes.

[0029] Step 106: Outcome Determination and Interpretability Output. The predicted probability value obtained in Step 105 is compared with a preset threshold, and the corresponding pregnancy outcome classification result is output. Simultaneously, during the output of the prediction results, the system generates interpretable visualization results based on the model's internal features and gradient information. Specifically, using a gradient-based class activation mapping method, key embryonic regions of interest to the model during prediction are highlighted, and the generated heatmap is overlaid on the original embryonic image to visually present the morphological feature regions upon which the model's decisions are based. Explanation of the effects of explanatory implementation

[0030] Through the aforementioned interpretable output methods, the model can not only provide predictions of pregnancy outcomes but also simultaneously output visual explanatory information corresponding to the predictions. In the single embryo transfer scenario, interpretable results are used to reveal the key embryonic structural regions that the model focuses on; in the two embryo transfer scenario, attention weights and heatmaps can jointly reflect the relative importance of different embryos in the prediction decision.

[0031] This interpretable output mechanism enhances the transparency of the model's prediction process, helps clinicians understand the basis of the model's decisions, and improves the credibility and acceptability of AI-assisted decision-making in the field of assisted reproduction. Technical effects of this embodiment

[0032] Through the above embodiments, the present invention realizes the automatic differentiation between clinical pregnancy and biochemical pregnancy outcomes based on static embryo images before embryo transfer, and provides intuitive and traceable interpretation results while ensuring prediction accuracy, providing reliable technical support for embryo selection and clinical decision-making.

Claims

1. A predictive method for distinguishing between clinical pregnancy and biochemical pregnancy outcomes based on static embryo images, characterized in that, Includes the following steps: 1) Obtain a static image of at least one embryo for which the transplant outcome is to be predicted; 2) Preprocess the static embryo images to obtain standardized image input data; 3) Input the image input data into the trained deep learning model, which is configured to extract multi-scale embryo morphological features from the image input data; when the input corresponds to double embryo transfer, the features from different embryos are adaptively weighted and fused through the attention fusion module; based on the fused feature vector, the predicted probability of the transfer cycle being a clinical pregnancy is output. 4) Generate pregnancy outcome prediction results based on the predicted probabilities to distinguish between clinical pregnancy and biochemical pregnancy.

2. The prediction method according to claim 1, characterized in that, In clinical applications, the method is implemented by a computer device executing instructions configured to perform the following steps: 1) Data input steps: Obtain one or two static images of embryos from the same patient; 2) Image preprocessing steps: background removal, normalization, and size standardization are performed on the embryo image; 3) Feature extraction and fusion steps: Input the preprocessed image into the deep learning model to extract embryo features, and fuse the features according to the transplantation scenario; 4) Outcome prediction step: Output the prediction results of clinical pregnancy or biochemical pregnancy based on the fused features.

3. The prediction method according to claim 1, characterized in that, The preprocessing steps include: 1) The original embryo image is segmented to remove the background area and retain the main body of the embryo; 2) Normalize and standardize the size of the segmented embryo images; 3) During the model training phase, data augmentation operations are applied to the training set images, including random rotation, random scaling and cropping, horizontal flipping, and color jitter.

4. The prediction method according to claim 1, characterized in that, The feature extraction module of the deep learning model includes a convolutional neural network backbone and a feature pyramid network, which are used to extract and fuse multi-scale feature information of embryo images.

5. The prediction method according to claim 1, characterized in that, The attention fusion module is configured to: receive feature vectors from different embryos; generate corresponding attention weights based on the feature vectors; and use the attention weights to perform weighted fusion of embryo features to characterize the relative contribution of different embryos to the prediction of pregnancy outcomes.

6. The prediction method according to claim 5, characterized in that: When the transfer cycle is a double embryo transfer, the attention fusion module performs adaptive weighted fusion of the features of the two embryos; when the transfer cycle is a single embryo transfer, the attention fusion module completes the fusion process based on the corresponding embryo features.

7. The prediction method according to claim 1, characterized in that, The method further includes an interpretability analysis step: generating a class activation map heatmap based on gradient information from a deep learning model; and using the heatmap to visualize the embryo image region that the model focuses on during the prediction process, in order to help understand the basis for pregnancy outcome prediction.

8. The prediction method according to claim 4, characterized in that, The feature pyramid network constructs a multi-scale feature representation for embryo morphological feature analysis through horizontal connections of multi-level features and top-down fusion.

9. The prediction method according to claim 5, characterized in that, The attention fusion module includes: at least one fully connected layer; an activation function; and a normalization function for generating normalized attention weights.

10. The prediction method according to claim 5, characterized in that, The attention fusion module calculates the fused feature vector in the following manner: ； ； in, and Given two feature vectors as input, This represents a vector concatenation operation. and These are the learnable weights and biases of the first and second fully connected layers, respectively. , For the generated attention weights, The final output is the fused feature vector.

Images