A deep learning-based non-invasive screening method for genetically edited rice resistant lines

Abstract

The application discloses a kind of gene editing rice resistance strain based on deep learning non-injury screening method, belong to plant genetic engineering field.The method is obtained to T0 generation or T1 generation regeneration plant population after Xa21 disease resistance gene is edited by CRISPR / Cas9 system, through the image of leaf after inoculation white leaf blight bacteria is collected, using the rice disease resistance gene editing effect detection model comprising ResNet50 encoder, multiscale context perception feature calibration module, morphological prior module, segmentation head and identification head, realize pixel level disease spot segmentation, six-dimensional morphological feature extraction and resistance grade discrimination under the premise of not destroying the integrity of plant, so as to quickly screen positive strain carrying effective editing event from regeneration plant population, replace existing gene editing positive strain screening, significantly improve the efficiency and flux of early screening of gene editing breeding.

Description

Technical Field

[0001] This application relates to the field of plant genetic engineering, specifically to a non-destructive screening method for gene-edited rice resistant lines based on deep learning. Background Technology

[0002] Rice bacterial blight, caused by Xanthomonas oryzae, is a major global disease that seriously threatens rice yield and quality. Targeted editing of the Xa21 resistance gene in rice using the CRISPR / Cas9 system is a crucial method for improving resistance to bacterial blight. After transforming the editing vector into rice callus tissue to obtain T0 generation regenerated plants, positive lines carrying valid editing events (defined by DNA sequencing as expected sequence changes at the target gene site, including but not limited to base insertions, deletions, or substitutions) need to be selected for subsequent breeding processes. However, the number of T0 generation plants is limited and their growth status is fragile. Traditional PCR molecular detection requires collecting leaf tissue for DNA extraction, which is destructive sampling and may lead to the death or growth stunting of positive plants, resulting in irreversible loss of gene-edited material.

[0003] Existing molecular detection methods based on PCR amplification, T7E1 restriction enzyme digestion, or Southern blot all suffer from drawbacks such as long processing times (2-3 days / plant), low throughput (20-30 plants / day), and irreversible damage, severely limiting the efficiency of early screening in gene editing breeding. Furthermore, local sampling of T0 generation chimeric plants may not represent the overall editing status of the plant, leading to missed detections or misjudgments.

[0004] Although some studies have attempted to use computer vision to assist in plant phenotypic analysis, existing methods are mostly designed for general disease identification and have not been optimized for gene editing scenarios: they do not consider the resistance phenotypic gradient caused by differences in editing efficiency, the local lesion heterogeneity caused by chimeras, and lack connection with the gene editing screening process, so they cannot be directly used for positive line screening decisions.

[0005] Therefore, there is an urgent need for a non-destructive, high-throughput early screening method for CRISPR / Cas9 gene-edited rice, which can quickly identify positive lines without damaging the plants at the T0 generation stage, replace traditional PCR detection, and improve the material preservation rate and screening efficiency of gene-edited breeding. Summary of the Invention

[0006] In view of this, this application provides a deep learning-based non-destructive screening method for gene-edited resistant rice lines. This method is used to rapidly and non-destructively evaluate the bacterial blight resistance phenotype of T0 or T1 generation regenerated plant populations obtained after editing the Xa21 disease resistance gene using the CRISPR / Cas9 system, in order to screen positive lines carrying effective editing events, replacing molecular detection methods based on PCR amplification. This method is implemented through a rice disease resistance gene editing effect detection model, which includes a ResNet50 encoder and a decoder, segmentation head, and discrimination head with morphological prior modules and multi-scale context-aware feature calibration modules. The method includes: T0 generation regenerated plants obtained by CRISPR / Cas9 vector transformation and differentiation culture were obtained. The T0 generation regenerated plants were inoculated with bacterial blight pathogen. After inoculation, images of rice leaves to be tested were collected. The preprocessed images of rice leaves to be tested were input into the ResNet50 encoder to obtain five coding feature maps at different scales. The decoder's multi-scale context-aware feature calibration module dynamically calibrates the encoded feature maps at different scales and generates a fourth-level decoded feature map based on the calibrated features. The decoder's morphological prior module extracts features from the second-level decoded initial feature map to obtain a six-dimensional morphological feature vector. The second-level decoded initial feature map is generated based on the third-level decoded feature map and the calibrated features of the encoded feature map at the third scale. The last level of the four-level decoding feature map is input into the segmentation head to obtain a pixel-level disease segmentation probability map; the six-dimensional morphological feature vector and the bottleneck feature map are input into the identification head to obtain a resistance level probability distribution map, and the category corresponding to the highest probability in the resistance level probability distribution map is taken as the target resistance level; based on the target resistance level, it is determined whether the T0 generation regenerated plant carries a valid Xa21 gene editing event; wherein, the bottleneck feature map is the feature map with the smallest spatial resolution and the most channels among multiple coding feature maps of different scales; The six-dimensional morphological feature vector, pixel-level disease segmentation probability map, and target resistance level of positive lines identified as carrying valid editing events are visualized and displayed. The positive lines are marked as candidate lines to enter the T1 generation homozygous screening, and plants identified as negative or invalid editing are removed from the screening process.

[0007] Optionally, the CRISPR / Cas9 vector comprises a gRNA expression cassette targeting the Xa21 gene and a Cas9 protein expression cassette, wherein the target sequence of the gRNA expression cassette is located in the coding region or promoter region of the Xa21 gene, and is used to induce loss-of-function mutations or changes in promoter activity of the Xa21 gene.

[0008] Optionally, the ResNet50 encoder includes five processing stages, each outputting an encoded feature map at one scale; the decoder includes four multi-scale context-aware feature calibration modules. The multi-scale context-aware feature calibration module of the decoder dynamically calibrates the encoded feature maps at different scales and generates a four-level decoded feature map based on the calibrated features, including: The i-th multi-scale context-aware feature calibration module concatenates the input feature map of the i-th upsampling stage and the encoded feature map of the i-th scale along the channel dimension, and generates the i-th spatial attention feature through a 3x3 convolutional layer and a sigmoid function. Global average pooling is then applied to the input feature map of the i-th upsampling stage and the encoded feature map of the i-th scale, and the concatenation is followed by a multilayer perceptron and a sigmoid function to generate the i-th channel attention feature. The i-th spatial attention feature and the i-th channel attention feature are then used to element-wise weight the encoded feature map of the i-th scale, and residual connections are performed to obtain the i-th calibrated feature. Based on the i-th calibrated feature and the input feature map of the i-th upsampling stage, the i-th level decoding feature map is generated. Here, i is an integer greater than 0 and less than 5.

[0009] Optionally, the ith spatial attention feature and the ith channel attention feature are used to perform element-wise weighting on the encoded feature map at the ith scale, and residual connection is performed to obtain the ith calibrated feature. This includes: reshaping the ith spatial attention feature and broadcasting it in the spatial dimension to obtain the ith reference spatial attention feature map; broadcasting the ith channel attention feature in the spatial dimension to obtain the ith reference channel attention feature map; multiplying the ith reference spatial attention feature map, the ith reference channel attention feature map, and the encoded feature map at the ith scale element-wise, and adding the encoded feature map at the ith scale as a residual term to obtain the ith calibrated feature.

[0010] Optionally, the process of generating the initial feature map for the second-level decoding includes: performing a 2x upsampling and 1x1 convolutional layer processing on the third-level decoding feature map to obtain a guiding feature map; the number of channels in the guiding feature map is the same as the number of channels in the encoding feature map at the third scale; dynamically calibrating the encoding feature map at the third scale using the third multi-scale context-aware feature calibration module to obtain the third calibrated feature; concatenating the guiding feature map and the third calibrated feature along the channel dimension, and processing them through two 3x3 convolutional layers to generate the initial feature map for the second-level decoding.

[0011] Optionally, the morphological prior module of the decoder is used to extract features from the initial feature map of the second-level decoding to obtain a six-dimensional morphological feature vector, including: The initial feature map of the second-level decoding is processed through a 1x1 convolutional layer using a morphological prior module to obtain a reference feature map. Dilation and erosion operations are then performed on the reference feature map to obtain dilated and eroded feature maps, respectively. A soft thresholding function is used to process the dilated and eroded feature maps to obtain dilated lesion probability maps and eroded lesion probability maps, respectively. Three morphological index values ​​are extracted from the dilated and eroded lesion probability maps using differentiable statistical methods. A six-dimensional morphological feature vector is constructed based on the extracted three morphological index values, including average lesion area, number of lesions, and lesion dispersion.

[0012] Optionally, the method further includes: adding the reference feature map, the dilated feature map, and the eroded feature map element by element to obtain an enhanced feature map; adjusting the number of channels of the enhanced feature map through a 1x1 convolutional layer to obtain the input feature map of the second upsampling stage; and generating a first-level decoded feature map based on the input feature map of the second upsampling stage, the second multi-scale context-aware feature calibration module, and the second-scale encoded feature map.

[0013] Optionally, the last level of the four-level decoded feature map is input into the segmentation head to obtain a pixel-level disease segmentation probability map, including: using the 1x1 convolution module in the segmentation head to perform convolution processing on the last level of the four-level decoded feature map to obtain a two-channel feature map, and then using the Softmax function in the segmentation head to generate a pixel-level disease segmentation probability map.

[0014] Optionally, the six-dimensional morphological feature vector and the bottleneck feature map are input into the discrimination head to obtain the resistance level probability distribution map, including: using the global average pooling module in the discrimination head to perform global average pooling on the bottleneck feature map to obtain a global feature vector; using the feature concatenation module in the discrimination head to concatenate the global feature vector and the six-dimensional morphological feature vector to obtain a comprehensive feature vector; and using the three fully connected layers and the Softmax function in the discrimination head to process the comprehensive feature vector to obtain the resistance level probability distribution map.

[0015] Optionally, the training process of the rice disease resistance gene editing effect detection model includes: Images of rice leaf samples treated with rice disease resistance gene editing were collected, and pixel-level binary mask images and resistance level labels annotated by experts were obtained. The rice leaf sample images were input into the initial rice disease resistance gene editing effect detection model to obtain pixel-level disease segmentation prediction probability maps and resistance level prediction probability distribution maps. Based on the resistance level corresponding to the maximum probability value in the pixel-level disease segmentation prediction probability maps and resistance level prediction probability distribution maps, the binary mask images, the resistance level labels, and the total loss function constructed from segmentation loss and identification loss, the network parameters of the initial rice disease resistance gene editing effect detection model were backpropagated and trained until the convergence condition was met, resulting in a trained rice disease resistance gene editing effect detection model.

[0016] In this embodiment, the preprocessed image of the rice leaf to be detected is input into a ResNet50 encoder to obtain five encoded feature maps at different scales. The multi-scale context-aware feature calibration module of the decoder dynamically calibrates the encoded feature maps at different scales and generates a four-level decoded feature map based on the calibrated features. The morphological prior module of the decoder extracts features from the initial second-level decoded feature map to obtain a six-dimensional morphological feature vector. The last-level decoded feature map in the four-level decoded feature map is input into the segmentation head to obtain a pixel-level disease segmentation probability map. The morphological feature vector and bottleneck feature map are input into the identification head to obtain the resistance level probability distribution map, and the category corresponding to the highest probability in the resistance level probability distribution map is taken as the target resistance level. Based on the target resistance level, it is determined whether the T0 generation regenerated plants carry a valid Xa21 gene editing event. The six-dimensional morphological feature vector, pixel-level disease segmentation probability map and target resistance level of the positive lines identified as carrying valid editing events are visualized and displayed. The positive lines are marked as candidate lines to enter the T1 generation homozygous screening, and the plants identified as negative or invalid editing are removed from the screening process. Thus, by replacing PCR testing with leaf image analysis, DNA extraction from leaf tissue is eliminated, avoiding destructive sampling of valuable T0 generation regenerated plants and improving screening throughput and efficiency. The multi-scale context-aware feature calibration module of the decoder dynamically calibrates the encoded feature maps at different scales, enhancing the encoder's sensitivity to subtle lesions. This allows the model to focus on truly disease-related areas, improving the accuracy of the pixel-level disease segmentation probability map output by the segmentation head. The morphological prior module of the decoder extracts features from the initial feature map of the second-level decoding, yielding biologically significant lesion morphological features, i.e., a six-dimensional morphological feature vector. Furthermore, the six-dimensional morphological feature vector and the bottleneck feature map are input into the identification head to obtain a resistance level probability distribution map. Using the category with the highest probability in the resistance level probability distribution map as the target resistance level provides accurate identification results. Furthermore, the six-dimensional morphological feature vector, pixel-level disease segmentation probability map, and target resistance level of positive lines identified as carrying valid editing events are visualized. Positive lines are marked as candidate lines to enter T1 generation homozygous screening, while plants identified as negative or invalid edits are removed from the screening process. This allows researchers to not only intuitively obtain resistance levels based on the visualized identification results, but also to obtain transparent and interpretable morphological evidence and decision-making suggestions, significantly improving the trust and adoption rate of the identification results. At the same time, it realizes an automated screening pipeline from gene editing transformation to resistance breeding. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A flowchart illustrating a non-destructive screening method for gene-edited rice resistant lines based on deep learning, provided for embodiments of this application; Figure 2 A schematic diagram of the architecture of a rice disease resistance gene editing effect detection model provided in an embodiment of this application; Figure 3 A schematic diagram of the data processing flow of a fourth multi-scale context-aware feature calibration module provided in this application embodiment; Figure 4 This is a schematic diagram of the data processing flow of a morphological prior module provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0020] It should be noted that the terms "first, second, third, and fourth" used in the embodiments of the present invention are only used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third, and fourth" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described herein can be implemented in an order other than that illustrated or described herein.

[0021] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0022] Non-destructive early screening: refers to a method for rapidly evaluating the editing events and resistance phenotypes of gene-edited plants through non-contact image acquisition and deep learning analysis without collecting plant tissues, extracting DNA, or performing destructive operations such as PCR amplification.

[0023] Editing events: refer to DNA double-strand breaks induced by the CRISPR / Cas9 system at target gene sites and subsequent insertions, deletions, or substitutions resulting from repair.

[0024] Positive lines: These are plants that have been identified through non-destructive screening as carrying valid editing events and exhibiting a disease-resistant phenotype (moderate or high resistance), and can proceed to the next breeding process.

[0025] T0 generation regenerated plants: refer to the first generation of transgenic / gene-edited plants obtained after transformation and differentiation culture by gene editing vectors, usually in a chimeric or heterozygous state.

[0026] In this embodiment, the CRISPR / Cas9 system, well-known in the art, is used to directionally edit the rice Xa21 disease resistance gene. Specific steps such as vector construction, Agrobacterium-mediated transformation, and differentiation culture are performed according to conventional operating procedures in the art and will not be elaborated here. The following focuses on describing the non-invasive early screening method of this application.

[0027] This application provides a non-destructive screening method for gene-edited resistant rice lines based on deep learning. This method is implemented through a rice disease resistance gene-editing effect detection model, which includes a ResNet50 encoder, a decoder with a morphological prior module and a multi-scale context-aware feature calibration module, a segmentation head, and a discrimination head. Figure 1 The diagram shown is a flowchart of a non-destructive screening method for gene-edited rice resistance lines based on deep learning, provided in this application. The method includes: S101. Obtain T0 generation regenerated plants transformed and differentiated by CRISPR / Cas9 vector, inoculate the T0 generation regenerated plants with bacterial blight pathogen, and collect images of rice leaves to be tested after inoculation; input the preprocessed images of rice leaves to be tested into the ResNet50 encoder to obtain five coding feature maps at different scales.

[0028] Here, rice plant samples were obtained after CRISPR / Cas9 gene editing. The rice plant samples were T0 generation regenerated plants obtained by transformation and differentiation culture using CRISPR / Cas9 vector. The CRISPR / Cas9 vector was used to target and edit the rice Xa21 disease resistance gene. The T0 generation regenerated plants were inoculated with bacterial blight pathogens, and images of the rice leaves to be tested were collected on the 7th day after inoculation. The T0 generation regenerated plants were obtained using vector construction and Agrobacterium-mediated transformation methods known in the art using the CRISPR / Cas9 system.

[0029] In some embodiments, the rice leaf image to be detected is a leaf image of rice that has undergone genetic editing to develop resistance genes for rice bacterial blight, acquired through an image acquisition device. Preprocessing the rice leaf image to be detected includes: processing the acquired rice leaf image... ( These represent the height, width, and number of channels of the rice leaf image to be detected, respectively. ) Scale to a fixed size The image is then normalized to obtain a standardized image. (i.e., the preprocessed image of the rice leaf to be detected).

[0030] In some embodiments, the ResNet50 encoder includes five processing stages, each outputting a encoded feature map at one scale, thereby obtaining five encoded feature maps at different scales. For example, as... Figure 2 As shown, the first processing stage (Stage 1) of the ResNet50 encoder outputs the encoded feature map at the first scale. ( The second processing stage (Stage 2) outputs the encoded feature map at the second scale. ( The third processing stage (Stage 3) outputs the encoded feature map at the third scale. ( The fourth processing stage (Stage 4) outputs the encoded feature map at the fourth scale. ( The fifth processing stage (Stage 5) outputs the encoded feature map at the fifth scale. ( ).

[0031] S102. The multi-scale context-aware feature calibration module of the decoder dynamically calibrates the encoded feature maps at different scales and generates a four-level decoded feature map based on the calibrated features. The morphological prior module of the decoder extracts features from the initial feature map of the second-level decoder to obtain a six-dimensional morphological feature vector.

[0032] The second-level decoding initial feature map is generated based on the third-level decoding feature map and the calibrated feature map of the third scale.

[0033] In some embodiments, such as Figure 2 As shown, the decoder includes four multi-scale context-aware feature calibration modules. The four-level decoding feature map includes a third-level decoding feature map, a second-level decoding feature map, a first-level decoding feature map, and a last-level decoding feature map. The third-level decoding feature map is the shallowest decoding feature map, and the last-level decoding feature map is the deepest decoding feature map, which is the final decoding feature map output by the decoder.

[0034] In some embodiments, the fourth multi-scale context-aware feature calibration module can dynamically calibrate the encoded feature map at the fourth scale and generate a third-level decoded feature map based on the calibrated features and the encoded feature map at the fifth scale; the third multi-scale context-aware feature calibration module can dynamically calibrate the encoded feature map at the third scale and generate a second-level decoded feature map based on the calibrated features and the third-level decoded feature map; the second multi-scale context-aware feature calibration module can dynamically calibrate the encoded feature map at the second scale and generate a first-level decoded feature map based on the calibrated features and the second-level decoded initial feature map; the first multi-scale context-aware feature calibration module can dynamically calibrate the encoded feature map at the first scale and generate a final-level decoded feature map based on the calibrated features and the first-level decoded feature map.

[0035] In some embodiments, the multi-scale context-aware feature calibration module of the decoder dynamically calibrates the encoded feature maps at different scales, and generates a four-level decoded feature map based on the calibrated features, including: The i-th multi-scale context-aware feature calibration module concatenates the input feature map of the i-th upsampling stage and the encoding feature map of the i-th scale along the channel dimension, and generates the i-th spatial attention feature through a 3x3 convolutional layer and a sigmoid function. Global average pooling is then applied to the input feature map of the i-th upsampling stage and the encoding feature map of the i-th scale, and the concatenation is followed by a multilayer perceptron and a sigmoid function to generate the i-th channel attention feature. The i-th spatial attention feature and the i-th channel attention feature are then used to element-wise weight the encoding feature map of the i-th scale, and residual connections are performed to obtain the i-th calibrated feature. Based on the i-th calibrated feature and the input feature map of the i-th upsampling stage, the (i-1)-th level decoding feature map is generated. Here, i is an integer greater than 0 and less than 5.

[0036] In some embodiments, the element-wise weighting of the encoded feature map at the i-th scale using the i-th spatial attention feature and the i-th channel attention feature, followed by residual connection, yields the i-th calibrated feature. This includes: reshaping the i-th spatial attention feature and broadcasting it in the spatial dimension to obtain the i-th reference spatial attention feature map; broadcasting the i-th channel attention feature in the spatial dimension to obtain the i-th reference channel attention feature map; and performing element-wise multiplication of the i-th reference spatial attention feature map, the i-th reference channel attention feature map, and the encoded feature map at the i-th scale, and adding the encoded feature map at the i-th scale as a residual term to obtain the i-th calibrated feature.

[0037] In some embodiments, when i=4, the input feature map of the fifth upsampling stage is obtained by performing a 2x upsampling and 1x1 convolutional layer processing on the encoded feature map of the fifth scale. Specifically, the encoded feature map of the fifth scale... The bottleneck feature map is upsampled by a factor of two using bilinear interpolation, and then processed through a 1x1 convolutional layer to adjust the number of channels to 1024, resulting in a feature map that matches the size of the fourth-level encoded feature space, i.e., the input feature map of the fourth upsampling stage. After that, and Input to the fourth multi-scale context-aware feature calibration module, and the fourth multi-scale context-aware feature calibration module performs... Figure 3 The data processing procedure shown includes: [The following is a list of steps, not a direct translation] and By stitching along the channel dimension, the fused features are obtained. The number of channels is compressed from 2048 to 1 (i.e., the number of output channels is 1) through a 3×3 convolutional layer, and then a fourth spatial attention map is generated by sigmoid activation. The 3×3 convolutional layer has a kernel size of 3×3, 2048 input channels, 1 output channel, and 1 padding layer, while maintaining the same spatial resolution.

[0038] At the same time, respectively and Perform global average pooling to obtain the feature vector. The concatenated vectors are input into a multilayer perceptron (MLP) and activated by a sigmoid function to generate a fourth channel attention vector. The MLP structure consists of a fully connected layer, ReLU activation, and another fully connected layer. The first fully connected layer reduces the dimension of the concatenated feature vector from 1024 to 256, while the second fully connected layer increases the dimension of the feature vector after ReLU activation from 256 to 1024.

[0039] Next, the fourth spatial attention map is used. and the fourth channel attention vector Encoded feature map for the fourth scale Element-wise weighted calibration is performed, followed by residual linking. Specifically, the following steps are taken: Reshape (reshape operation) The shape is then broadcast in the spatial dimension to obtain a fourth reference space attention feature map. At the same time, Broadcasting along the channel dimension yields the fourth reference channel attention feature map. Where B is the batch size, i.e., the number of samples processed in one forward propagation. The value of B is determined by the number of input data. During the training phase, B is usually set to positive integers such as 8, 16, and 32 to fully utilize parallel computing capabilities and stabilize gradient estimation. During the testing or online application phase, B is set to 1 (processing one image at a time) or other positive integers smaller than the training batch size (in this application, B=1 is used for the online application phase). The network structure has adaptive capabilities to B, and there is no need to modify the model parameters.

[0040] Then, and and Perform element-wise multiplication and... Added as residual terms, we obtain the fourth calibrated feature. ,in, This indicates element-wise multiplication.

[0041] After that, and By stitching along the channel dimension, the stitching feature is obtained. Upsampled by 2 times using bilinear interpolation The upsampled feature map is obtained. Then after two Convolutional layers (each followed by batch normalization and ReLU activation function) are used for feature fusion and channel dimensionality reduction. The first convolutional layer reduces the number of channels from 2048 to 512, and the second convolutional layer reduces the number of channels from 512 to 256, ultimately generating the third-level decoded feature map. .

[0042] In some embodiments, the process of generating the initial feature map for the second-level decoding includes: performing a 2x upsampling and 1x1 convolutional layer processing on the third-level decoding feature map to obtain a guiding feature map; the number of channels in the guiding feature map is the same as the number of channels in the encoded feature map at the third scale; dynamically calibrating the encoded feature map at the third scale using a third multi-scale context-aware feature calibration module to obtain the third calibrated feature; concatenating the guiding feature map and the third calibrated feature along the channel dimension, and processing them through two 3x3 convolutional layers to generate the initial feature map for the second-level decoding.

[0043] In some embodiments, when i=3, for Perform bilinear interpolation upsampling twice to resolution After The number of channels in the convolutional layer is adjusted to 512 to obtain the input feature map for the third upsampling stage. (Guidance feature map), then, and Before stitching along the channel dimension, Perform bilinear interpolation upsampling by a factor of two to The upsampled feature map is obtained. To ensure that the two dimensions are consistent; and The data is concatenated along the channel dimension, and a third spatial attention feature map is generated by passing it through a 3x3 convolutional layer and a sigmoid function. At the same time, respectively for and Global average pooling is performed, and a third channel attention vector is generated using MLP and Sigmoid activation. Then, The image is reshaped and broadcast in the spatial dimension to obtain a third reference space attention feature map. At the same time, Broadcasting along the channel dimension yields a third reference channel attention feature map. Then, and and Perform element-wise multiplication and... Added as residual terms, we obtain the third calibrated feature. After that, and spliced ​​along the channel dimension, through two The convolutional layers (with channel dimensions decreasing sequentially from 512 to 256, and then from 256 to 128) perform feature fusion and channel dimensionality reduction to generate second-level decoding features. ,here These are the initial decoding features for the second level, which need to be processed by the subsequent morphological prior module before further decoding.

[0044] In some embodiments, a six-dimensional morphological feature vector is obtained by extracting features from the initial feature map of the second-level decoding using the morphological prior module of the decoder. This includes: processing the initial feature map of the second-level decoding using a 1x1 convolutional layer to obtain a reference feature map; performing dilation and erosion operations on the reference feature map to obtain a dilated feature map and an eroded feature map; processing the dilated and eroded feature maps using a soft thresholding function to obtain a dilated lesion probability map and an eroded lesion probability map; extracting three morphological index values ​​from the dilated and eroded lesion probability maps using differentiable statistical methods; and constructing a six-dimensional morphological feature vector based on the extracted three morphological index values. The three morphological index values ​​include average lesion area, number of lesions, and lesion dispersion.

[0045] In some embodiments, such as Figure 4 The diagram shown illustrates the data processing flow of a morphological prior module provided in this application. First, the second-level initial decoded features are processed through a 1x1 convolutional layer. The number of channels is compressed to 64 to obtain the reference feature map. For reference feature map Differentiable expansion and erosion operations are performed separately to obtain expansion feature maps. Corrosion feature diagram This application employs learnable structural elements. Differentiability is achieved through a smooth approximation, and the dilation operation is expressed by formula (1): (1); in, These represent the height and width coordinates of a pixel in the feature map, respectively. Indicates the channel index. These represent the offset in the height direction and the offset in the width direction, respectively. , Indicates the location of the channel in the feature map ,Location The feature value or pixel value, Represents structural element Middle position elements, For temperature coefficient, this application takes .

[0046] The corrosion operation is represented by formula (2): (2); Then, the soft threshold function shown in formula (3) is used. dilation feature map Corrosion feature diagram The probabilities of the swelling lesions were obtained by processing them separately. Probability diagram of corrosion spots .

[0047] (3); in, Input feature values ​​(pixel values ​​of pixels in the dilated feature map or the feature map itself). The threshold parameter is a learnable parameter (initially 0.5). The learnable dimension coefficient (initial value is 10).

[0048] Then, targeting and Differentiable statistical methods (such as connected component analysis) were used to extract three morphological indices: average lesion area, lesion number, and lesion dispersion. Among these, the average lesion area... (The expected proportion of the diseased area to the total leaf area) is calculated using formula (4): (4); in, Probability map of swelling lesions Or corrosion lesion probability diagram The middle pixel coordinates are The probability value of each pixel.

[0049] The number of lesions is based on the lesion probability map (probability map of inflated lesions). Or corrosion lesion probability diagram Local maxima detection and soft counting are performed by first applying Gaussian smoothing. , The standard deviation of the Gaussian kernel is represented. The pixel coordinates in the Gaussian smoothed lesion probability map are ( The probability value of the pixel is calculated; then the soft local maximum is calculated using formula (5). Finally, the number of lesions is obtained by summing using formula (6). : (5); in, The pixel coordinates are ( The soft contribution of a pixel to the number of lesions in its local neighborhood (e.g., a 3x3 range) is called a soft local maximum. The pixel coordinates in the Gaussian smoothed lesion probability map are ( The probability value of the pixel whose coordinates are ( ). The pixel coordinates of ) are ( The pixels within the local neighborhood of a pixel (e.g., a 3x3 range, the size of which can be adaptively adjusted according to the size of the lesion probability map).

[0050] (6); in, The pixel coordinates are ( The pixel's soft contribution to the number of lesions in its local neighborhood is calculated. Lesion dispersion is calculated based on the second moment of the probability distribution, first by calculating the probability mass center: Then, the square root of the sum of variances is calculated using formula (7) as the lesion dispersion. : (7); Based on the probability map of dilated lesions The first average lesion area was calculated. Number of first lesions First lesion dispersion Based on the probability map of corrosion spots The second average lesion area was calculated. Number of second lesions Second lesion dispersion Afterwards, based on and Constructing a six-dimensional morphological feature vector: , .

[0051] In some embodiments, after obtaining the dilated feature map and the erosion feature map, the reference feature map, the dilated feature map, and the erosion feature map can be added element-wise to obtain the enhanced feature map; the number of channels of the enhanced feature map is adjusted by a 1x1 convolutional layer to obtain the input feature map of the second upsampling stage; based on the input feature map of the second upsampling stage, the second multi-scale context-aware feature calibration module, and the encoded feature map of the second scale, the first-level decoded feature map is generated.

[0052] In some embodiments, the reference feature map can be Dilation feature map Corrosion feature diagram Element-wise addition yields the enhanced feature map. Then, through a 1x1 convolutional layer... The number of channels is restored to 128, resulting in the input feature map for the second upsampling stage. , Replace the original Continue to participate in the subsequent decoding process.

[0053] The process of generating the first-level decoded feature map (i=2 at this time) based on the input feature map from the second upsampling stage, the second multi-scale context-aware feature calibration module, and the second-scale encoded feature map includes: and By splicing along the channel dimension, we obtain ,pass Convolution (384 input channels, 1 output channel) and the sigmoid function generate a second spatial attention map. At the same time, respectively for and Global average pooling is performed, and a second channel attention vector is generated using MLP and Sigmoid activation. .Will Remodeling The shape is then broadcast in the spatial dimension to obtain a second reference attention feature map. At the same time, Broadcast along the channel dimension to obtain a second reference channel attention feature map. Calculate the features after the second calibration. ,Will and splicing and through Convolution processing generates the first-level decoded feature map. .

[0054] In some embodiments, when i=1, the input feature map of the second upsampling stage is... Upsampled by 2 times to resolution via bilinear interpolation After The number of channels in the convolutional layer is adjusted to 64 to obtain the same features as the first-level features of the encoder. Guiding features for matching spatial size and number of channels (Input feature map of the first upsampling stage), using right The first calibrated feature is obtained by performing multi-scale context-aware feature calibration. (This process is similar to the multi-scale context-aware feature calibration process for i=2, 3, 4 mentioned above, and will not be described again.) and spliced ​​along the channel dimension, through two Convolutional layer (number of channels) The features are fused and refined, and then upsampled by a factor of 2 using bilinear interpolation to obtain the final decoded output feature map. (Final level decoded feature map).

[0055] S103. Input the last level of the decoding feature map in the four-level decoding feature map into the segmentation head to obtain a pixel-level disease segmentation probability map; input the six-dimensional morphological feature vector and the bottleneck feature map into the identification head to obtain a resistance level probability distribution map, and take the category corresponding to the highest probability in the resistance level probability distribution map as the target resistance level. Based on the target resistance level, determine whether the T0 generation regenerated plant carries a valid Xa21 gene editing event.

[0056] Among them, the bottleneck feature map is the feature map with the smallest spatial resolution and the most channels among multiple encoded feature maps of different scales.

[0057] Among them, plants with a target resistance level of "high resistance" or "medium resistance" were identified as positive lines carrying valid editing events, while plants with a target resistance level of "susceptible" were identified as negative lines that did not carry valid editing events or whose editing was invalid.

[0058] In some embodiments, the last level of the four-level decoded feature map is input into the segmentation head to obtain a pixel-level disease segmentation probability map, including: using the 1x1 convolution module in the segmentation head to perform convolution processing on the last level of the four-level decoded feature map to obtain a two-channel feature map, and then using the Softmax function in the segmentation head to generate a pixel-level disease segmentation probability map.

[0059] In some embodiments, the segmentation head processes the final-level decoded feature map of the decoder using a 1x1 convolutional module to obtain a two-channel feature map, which is then processed by the Softmax function to generate a pixel-level disease segmentation probability map. And obtain the binary segmentation map through the argmax operation. , and These represent pixel coordinates respectively. The probability that a pixel belongs to a diseased area or the background.

[0060] In some embodiments, the six-dimensional morphological feature vector and the bottleneck feature map are input into the discrimination head to obtain the resistance level probability distribution map, including: using the global average pooling module in the discrimination head to perform global average pooling on the bottleneck feature map to obtain a global feature vector; using the feature concatenation module in the discrimination head to concatenate the global feature vector and the six-dimensional morphological feature vector to obtain a comprehensive feature vector; and using the three fully connected layers and the Softmax function in the discrimination head to process the comprehensive feature vector to obtain the resistance level probability distribution map.

[0061] In some embodiments, the global average pooling module in the discrimination head will use the bottleneck feature map Perform global average pooling to obtain the global feature vector. global feature vector The global semantic information of the entire image is encoded. Then, the global feature vector is processed by the feature concatenation module in the discrimination head. The six-dimensional morphological feature vector extracted by the aforementioned morphological prior module By splicing along the channels, a comprehensive feature vector is obtained. Then, synthesize the feature vectors. The two fully connected layers in the discrimination head (with 512 and 128 neurons respectively, each followed by a ReLU activation and Dropout layer with a dropout ratio of 0.3) are sequentially input. Finally, the mixture is processed by a fully connected layer containing 3 neurons in the discrimination head and a Softmax function to obtain the resistance level probability distribution map. And the probability distribution map of resistance levels The category corresponding to the highest probability value is used as the target resistance level. In this system, 0 indicates infection, 1 indicates moderate resistance, and 2 indicates high resistance.

[0062] S104. Visualize the six-dimensional morphological feature vector, pixel-level disease segmentation probability map, and target resistance level of positive lines identified as carrying valid editing events, and mark the positive lines as candidate lines to enter the T1 generation homozygous screening, and remove the plants identified as negative or invalid edits from the screening process.

[0063] In this step, the screening results directly output a decision suggestion of "entering T1 generation homozygous screening" or "removal", realizing an automated screening pipeline from gene editing transformation to resistance breeding.

[0064] In some embodiments, the six-dimensional morphological feature vector, pixel-level disease segmentation probability map, and target resistance level can be used as the final identification result for visualization. The binary segmentation map can also be used. By overlaying the image onto the original image and color-coding diseased areas, a visually intuitive identification result map is generated, allowing researchers to clearly observe the distribution and morphology of the disease. Simultaneously, quantitative indicators are output, including the target resistance level. The six-dimensional morphological feature vector extracted by the morphological prior module is then used. Specific values ​​(including average lesion area) Number of lesions Lesion dispersion It outputs structured data, thus providing researchers with quantitative morphological analysis data.

[0065] In some embodiments of this application, the training process of the rice disease resistance gene editing effect detection model includes: acquiring rice leaf sample images processed by rice disease resistance gene editing, obtaining pixel-level binary mask images and resistance level labels annotated by experts; inputting the rice leaf sample images into the initial rice disease resistance gene editing effect detection model to obtain pixel-level disease segmentation prediction probability maps and resistance level prediction probability distribution maps; based on the resistance level corresponding to the maximum probability value in the pixel-level disease segmentation prediction probability maps and resistance level prediction probability distribution maps, the binary mask images, the resistance level labels, and the total loss function constructed from segmentation loss and identification loss, backpropagation training is performed on the network parameters of the initial rice disease resistance gene editing effect detection model until the convergence condition is met, thus obtaining the trained rice disease resistance gene editing effect detection model.

[0066] In some embodiments, a large number of original rice leaf images that have undergone rice disease resistance gene editing are collected. For these original rice leaf images, on the one hand, plant pathologists manually annotate the disease areas to generate pixel-level binary mask images. In this context, a pixel value of 1 represents a diseased area, and 0 represents a healthy area. and These represent the image height and width, respectively; on the other hand, based on industry standards such as the percentage of leaf area covered by lesions, experts comprehensively evaluate the resistance level of each sample. In this context, 0 represents susceptible, 1 represents moderately resistant, and 2 represents highly resistant. The labeled dataset is randomly divided into training, validation, and test sets in a 6:2:2 ratio.

[0067] Next, for each original image in the training set... ( (Number of channels: 3 for RGB images) Perform preprocessing to scale the image to a fixed size. ,get Subsequently, data augmentation operations were dynamically applied during training, including random rotation, random horizontal flipping, random brightness and contrast adjustment, and random cutout. Finally, the pixel values ​​were divided by 255 to map to the [0,1] interval, and standardized using the mean and standard deviation of the dataset to obtain standardized rice leaf sample images. .

[0068] Next, rice leaf sample images were input into the initial rice disease resistance gene editing effect detection model to obtain pixel-level disease segmentation prediction probability maps and resistance level prediction probability distribution maps. Based on the resistance level corresponding to the maximum probability value in the pixel-level disease segmentation prediction probability maps and resistance level prediction probability distribution maps, binary mask images, resistance level labels, and the total loss function constructed from segmentation loss and identification loss, the loss value was calculated, and the network parameters of the initial rice disease resistance gene editing effect detection model were backpropagated and trained using this loss value. During the training process, the Adam optimizer was used for parameter optimization, with an initial learning rate set to... The weight decay coefficient is A cosine annealing strategy is used to dynamically adjust the learning rate, where the maximum learning rate is... Minimum learning rate Total number of training rounds (Convergence criteria), batch size during training is set to... After each training epoch, the model performance is evaluated on the validation set, and the model parameters with the lowest loss on the validation set are saved as the final model.

[0069] In some embodiments, the total loss function is composed of a weighted average of segmentation loss and identification loss. This can be expressed by formula (8): (8); The weight hyperparameter is set to , This is to balance the optimization contributions of the two tasks. This is used to measure the accuracy of disease area segmentation, representing the loss from segmentation. To assess the damage and measure the accuracy of resistance classification, This can be expressed by formula (9): (9); in, Cross-entropy loss is used to measure the accuracy of pixel-level classification. The Dice loss measures the overlap between the predicted segmented region and the real region, and the combination coefficients are used. The weights used to balance the Dice loss and the cross-entropy loss; This can be expressed by formula (10): (10); in, For the spatial coordinate index of the pixel, This represents the disease probability predicted by the model. The actual binary mask labeled by the experts (the diseased area is 1). Used to avoid division by zero.

[0070] This can be expressed by formula (11): (11); in, These represent the height and width of the pixel-level disease segmentation probability map, respectively. For pixels Category one-hot real labels, Represents the pixel points predicted by the model. Category The probability, For category indexing, Indicates background, Indicates the area affected by the disease.

[0071] Loss assessment Different methods are used depending on the task type. For resistance level classification tasks, cross-entropy loss and identification loss are used. Represented as: (12); in, One-hot encoding representing the true resistance level, when the true level of the sample is hour The rest are 0; This indicates that the sample predicted by the model belongs to the resistance level. The probability value satisfies .

[0072] In some embodiments, the method further includes: collecting seeds from T0 generation regenerated plants identified as positive lines after T1 generation seeds mature for PCR-RFLP verification, and calculating the screening accuracy of the non-destructive early screening method; the screening accuracy is the proportion of positive lines identified by non-destructive screening that are indeed carrying the target editing event as verified by PCR. This demonstrates the reliability of non-destructive screening through comparison with traditional PCR detection.

[0073] In this embodiment, rice plant samples treated with CRISPR / Cas9 gene editing are first obtained: the rice plant samples are T0 generation regenerated plants obtained by CRISPR / Cas9 vector transformation and differentiation culture, and the CRISPR / Cas9 vector is used to target and edit the rice Xa21 disease resistance gene; the T0 generation regenerated plants are inoculated with bacterial blight pathogen, and leaf images are collected on the 7th day after inoculation; then, non-destructive phenotypic feature extraction is performed: the preprocessed rice leaf images to be tested are input into the ResNet50 encoder to obtain five encoded feature maps at different scales; the encoded feature maps at different scales are dynamically calibrated by the multi-scale context-aware feature calibration module of the decoder, and a four-level decoded feature map is generated based on the calibrated features; the morphological prior module of the decoder is used to extract features from the initial feature map of the second-level decoder. The process involves extracting a six-dimensional morphological feature vector; then, editing event discrimination and resistance level screening are performed: the last level of the four-level decoding feature map is input into the segmentation head to obtain a pixel-level disease segmentation probability map; the six-dimensional morphological feature vector and bottleneck feature map are input into the identification head to obtain a resistance level probability distribution map, and the category corresponding to the highest probability in the resistance level probability distribution map is taken as the target resistance level; based on the target resistance level, it is determined whether the T0 generation regenerated plants carry a valid Xa21 gene editing event; finally, positive lines are screened and visualized: the six-dimensional morphological feature vector, pixel-level disease segmentation probability map, and target resistance level of the positive lines identified as carrying valid editing events are visualized, and the positive lines are marked as candidate lines to enter the T1 generation homozygous screening, while plants identified as negative or invalid editing are removed from the screening process. In this way, leaf image analysis replaces PCR detection, eliminating the need for collecting leaf tissue for DNA extraction, avoiding destructive sampling of valuable T0 generation regenerated plants, and improving screening throughput and efficiency. Specifically, the multi-scale context-aware feature calibration module of the decoder dynamically calibrates the encoded feature maps at different scales, which enhances the encoder's sensitivity to subtle lesions, allowing the model to focus on areas truly related to the disease and improving the accuracy of the pixel-level disease segmentation probability map output by the segmentation head. The morphological prior module of the decoder extracts features from the initial feature map of the second-level decoding, which can extract biologically significant lesion morphological features, namely a six-dimensional morphological feature vector. Furthermore, the six-dimensional morphological feature vector and the bottleneck feature map are input into the identification head to obtain a resistance level probability distribution map. The category corresponding to the highest probability in the resistance level probability distribution map is taken as the target resistance level, which can obtain accurate identification results.Furthermore, the six-dimensional morphological feature vector, pixel-level disease segmentation probability map, and target resistance level of positive lines identified as carrying valid editing events are visualized. Positive lines are marked as candidate lines to enter T1 generation homozygous screening, while plants identified as negative or invalid edits are removed from the screening process. This allows researchers to not only intuitively obtain resistance levels based on the visualized identification results, but also to obtain transparent and interpretable morphological evidence and decision-making suggestions, significantly improving the trust and adoption rate of the identification results. At the same time, it realizes an automated screening pipeline from gene editing transformation to resistance breeding.

[0074] The following experiments illustrate the effectiveness of the method provided in this application. This experiment strictly adheres to standards in plant phenotyping and agricultural artificial intelligence, constructing a high-fidelity testing environment to verify the performance advantages of the proposed scheme in the identification of resistance in gene-edited rice. The experimental platform utilizes a self-developed high-throughput plant phenotyping system, including an industrial-grade RGB camera (resolution 5472×3648), a standard light box (light intensity stable at 5000 Lux, color temperature 6500K), and an automatic transport device to ensure consistent image acquisition conditions. The experimental dataset contains 1200 rice leaf images from 500 different gene-edited lines (CRISPR / Cas9 editing of the Xa21 gene). Each image was jointly annotated by three plant pathologists, with annotations including pixel-level disease area masks and resistance level labels (susceptible, moderately resistant, highly resistant). The dataset was divided into training, validation, and test sets in a 6:2:2 ratio to ensure the objectivity of the evaluation results.

[0075] Precise definition of disease type: Images were collected 7 days after bacterial blight inoculation. The lesion characteristics covered the entire disease course from early punctate infection to late streak fusion, including typical samples of different resistance phenotypes (susceptible: lesions spread rapidly, area greater than 30%; moderately resistant: lesions spread limited, area 10%~30%; highly resistant: allergic necrotic spots or no obvious lesions). The image acquisition environment simulated a real breeding scenario, including three types of changing factors: (1) light changes (4500~6500 Lux); (2) leaf posture changes (-15° to +15° tilt); (3) background interference (including soil, water stains, old leaves, etc.) to verify the robustness of the model in practical applications.

[0076] The comparison methods include: Baseline A (traditional image processing): based on threshold segmentation (Otsu method) and manual feature extraction (lesion area, length), representing the basic level of traditional identification methods; Baseline B (standard U-Net): a classic semantic segmentation network that only outputs lesion segmentation maps, and maps resistance levels by calculating lesion area ratios, reflecting the performance of pure segmentation methods; Baseline C (U-Net and classification head): adding a global classification branch on the basis of U-Net, adopting multi-task learning (segmentation and classification), but without feature calibration and morphological priors; The proposed solution: integrates a multi-scale context-aware feature calibration module and a morphological prior module.

[0077] The evaluation metrics comprehensively cover identification performance, including: segmentation accuracy: Dice coefficient (reflecting the accuracy of disease area segmentation); classification accuracy: Top-1 accuracy of resistance level determination (reflecting the reliability of identification results); interpretability metrics: correlation between morphological features (number of lesions, average area) and expert annotations (Pearson correlation coefficient); processing efficiency: average inference time per image (ms) and GPU memory usage (MB), ensuring complete verification from performance to engineering implementation.

[0078] Table 1 shows the performance comparison results of different methods in the task of identifying the effect of rice bacterial blight resistance gene editing: Table 1. Performance comparison of different methods in the task of identifying the effect of gene editing on resistance to bacterial blight in rice. ; As shown in Table 1, in terms of segmentation accuracy, the Dice coefficient of this application reaches 89.2%, which is 5.7 percentage points higher than the best-performing Baseline C (83.5%). In terms of resistance grade classification accuracy, this application achieves 93.8%, which is 9.2 percentage points higher than Baseline C (84.6%), showing a particularly significant improvement. This application's solution is far superior to the three baseline methods (Baseline A, Baseline B, and Baseline C) in terms of lesion number correlation coefficient (0.86) and lesion area correlation coefficient (0.91), with the highest being only 0.42 and 0.58 respectively. This fully demonstrates that the morphological feature vector extracted by the morphological prior module in this application has high accuracy and interpretability, and can accurately quantify the geometric attributes of lesions. In terms of inference efficiency, the inference time of this application is 52ms, which is slightly higher than Baseline C (48ms), but still meets the real-time requirements of practical application scenarios. The memory usage of this application is 1080MB, which is within an acceptable range. In summary, the proposed solution achieves significant improvements in segmentation accuracy, classification accuracy, and morphological feature consistency with relatively low computational overhead, verifying the effectiveness of the combined use of the multi-scale context-aware feature calibration module and the morphological prior module.

[0079] Furthermore, to clarify the independent contribution of each innovation, this experiment designed a rigorous ablation test: based on the application scheme (complete model), key components were removed one by one and performance degradation was measured.

[0080] Morphological prior module (removed and used directly with original features) The lack of morphological feature extraction led to a drop in classification accuracy to 87.3% (-6.5%) and a sharp drop in the correlation coefficient of lesion number to 0.42 (-51%), demonstrating that the morphological prior module enables the model to explicitly learn lesion morphology rather than implicitly fitting black-box features.

[0081] The multi-scale context-aware feature calibration module (which, after removal, uses standard U-Net skip connections without dynamic calibration) caused the Dice coefficient to drop to 84.1% (-5.1%) and the classification accuracy to drop to 88.5% (-5.3%). This demonstrates the core value of the multi-scale context-aware calibration module for identifying subtle lesions and cross-scale feature fusion.

[0082] In addition, to clarify the impact of each innovative module on the model inference efficiency, this experiment measured the average inference time (ms) and GPU memory usage (MB) of a single image under different model configurations in the same hardware environment (NVIDIA Tesla V100 32GB). All measurements were performed with a batch size of 1, and the average value of 100 inferences was taken.

[0083] Table 2 Comparison of performance parameters for inference efficiency of each innovation module ; As can be seen from Table 2, the addition of the multi-scale context-aware feature calibration module increases the inference time by about 5ms and the video memory by 120MB. The main overhead comes from the generation of multiple spatial and channel attention and element-wise weighted operations. However, the multi-scale context-aware feature calibration module contributes the most to the improvement of segmentation accuracy (Dice coefficient is improved by 5.1%).

[0084] The addition of the morphological prior module increases inference time by about 6ms and increases video memory by 140MB. The main overhead comes from the dilation operation, erosion operation, connected component analysis and forward computation of the soft threshold function. However, the morphological prior module contributes the most to the improvement of interpretability (the correlation of morphological features is improved by 0.44~0.46).

[0085] The complete model inference time is 52ms, which meets the real-time requirements of high-throughput plant phenotyping scenarios (greater than 15 frames / second). The video memory usage is 1080MB, which can be efficiently deployed on mainstream industrial-grade GPUs (such as NVIDIA RTX 3090 / 4090, with 24GB of video memory).

[0086] The inference time overhead of the two core modules (multi-scale context-aware feature calibration module and morphological prior module) is within a controllable range (a total increase of 11 ms), and the gains in accuracy and interpretability are significantly better than the increase in time cost, demonstrating good engineering feasibility in actual breeding and screening scenarios.

[0087] To verify the robustness of the model in real-world breeding scenarios, three sets of generalization tests were designed: Cross-variety testing: The model performance was tested on three rice varieties ("Minghui 63", "Teqing", and "Xiangwanxian 13") that had never been used in training. As the differences between varieties increased, the advantages of the proposed scheme were significantly maintained (91.2% accuracy on "Minghui 63", an improvement of 8.3% over Baseline C; and 88.7% accuracy on "Xiangwanxian 13", an improvement of 9.6% over Baseline C), which demonstrates the cross-variety adaptability of the multi-scale context-aware feature calibration module.

[0088] Cross-disease stress test: at inoculation concentration gradients (10) 7 ~10 9 Model stability was tested at CFU / mL. Low concentrations (10) 7 The accuracy rate of this application scheme is 86.5% (Baseline C is 72.3%), and the accuracy rate of high concentration (10) is [missing information]. 9The accuracy rate of the proposed solution is 94.2% (Baseline C is 86.8%), indicating that the proposed solution can maintain its advantages under different disease stresses, and is especially valuable in low-concentration early detection scenarios.

[0089] Illumination robustness test: Model performance fluctuations were tested within an illumination intensity range of 4500–6500 Lux. The standard deviation of the Dice coefficient of the proposed scheme was 1.8%, significantly lower than the 4.3% of Baseline C, demonstrating that the dynamic calibration mechanism of the multi-scale context-aware feature calibration module can effectively suppress feature perturbations introduced by illumination changes.

[0090] Under extreme conditions (low light 4500 Lux, high background interference, early lesions), the Dice coefficient of the proposed scheme (76.5%) was 31.5% higher than that of Baseline C (58.2%), effectively avoiding identification failure and fully verifying the engineering robustness of the proposed scheme in real complex breeding scenarios.

[0091] To verify the superiority of the morphological prior module in this application compared to existing differentiable morphological methods, this experiment uses the comparative method SoftMorph (a classic differentiable morphological network module that employs learnable structuring elements and differentiable approximations of morphological operations, commonly used for morphological prior embedding in image segmentation tasks). The comparative experiments were conducted with identical datasets and training settings, maintaining consistency in the encoder, multi-scale context-aware feature calibration module, segmentation head, identification head, and other hyperparameters. Only the morphological prior module was replaced with SoftMorph. The differences in segmentation accuracy, classification accuracy, morphological feature interpretability, and inference efficiency were compared between the two methods. The experimental results for different methods are shown in Table 3. Table 3 Comparison of experimental results from different methods ; As shown in Table 3, the morphological prior module adopted in this application improves classification accuracy by 4.6% and Dice coefficient by 2.8%. Simultaneously, the correlation between morphological features and expert annotations is significantly improved (the correlation coefficient for lesion number increases from 0.71 to 0.86, and the correlation coefficient for lesion area increases from 0.78 to 0.91). The inference time increases slightly (3 ms), but this is traded for higher interpretability and identification accuracy. Experimental results show that the morphological prior module, through decoupling dilation and erosion branches, adaptive binarization using a soft threshold function, and joint extraction of multiple indicators, can more accurately capture the morphological features of lesions while maintaining end-to-end training capabilities, outperforming general differentiable morphological methods.

[0092] When the T0 generation regenerated plants reached the 4-5 leaf stage, they were artificially inoculated with Xanthomonas oryzae pv. oryzae (strain PXO99A, concentration 10). 8 (CFU / mL). On day 7 post-inoculation, leaf images were acquired using an industrial-grade RGB camera under a standard light box; no leaf tissue sampling was required. The acquired leaf images were input into the rice disease resistance gene editing effect detection model provided in this application, and the resistance level (susceptible / moderately resistant / highly resistant) was output. Plants identified as "highly resistant" or "moderately resistant" were marked as positive candidate lines and directly entered the T1 generation for planting; plants identified as "susceptible" were removed from the screening process.

[0093] After the T1 generation seeds matured, the progeny of positive candidate lines were verified by PCR-RFLP. The proportion of positive lines identified by the non-destructive screening protocol of this application that actually carried the target editing event was calculated as the screening accuracy. The screening accuracy of the non-destructive screening protocol of this application reached 93.8%, and the concordance rate with traditional PCR detection was 91.2%, as shown in Table 4 below.

[0094] Table 4 Comparison of the non-destructive screening protocol and PCR detection in this application: ; The experimental results in Table 4 show that the non-destructive screening scheme of this application achieves non-contact, high-throughput, and rapid screening for resistance identification of gene-edited rice plants while maintaining a high accuracy rate of 93.8% (with a concordance rate of 91.2% with traditional PCR-RFLP detection). Compared with the limitations of traditional PCR-RFLP methods, which require destructive sampling, take 2-3 days per plant, and can only process 20-30 plants per batch, the non-destructive screening scheme of this application uses an industrial-grade RGB camera to acquire leaf images under a standard light box. After inputting the images into the detection model, the resistance level determination result can be output in 52ms. The throughput per batch exceeds 1000 plants / hour, and it causes no damage to the plants. Positive candidate lines can be directly planted in the T1 generation. The experimental results show that the non-destructive screening scheme of this application effectively solves the technical bottlenecks of low throughput, long cycle, and the need for destructive sampling in traditional PCR-RFLP methods. It has significant efficiency advantages and good practical value in the early large-scale screening application of gene-edited breeding materials.

[0095] Understandably, this application dynamically enhances the encoder's sensitivity to subtle lesions through a multi-scale context-aware feature calibration module, explicitly extracts biologically significant lesion morphological features through a morphological prior module, and outputs the resistance level after fusing global features and morphological features through joint optimization. This enables the identification system to not only accurately identify resistance but also provide transparent and interpretable morphological evidence, significantly improving the trustworthiness and acceptance rate of the identification results.

[0096] It should be noted that, depending on the implementation needs, the steps described in the embodiments of this application can be broken down into more steps, or two or more steps or parts of the steps can be combined into new steps to achieve the purpose of the embodiments of this application.

[0097] The methods described in the embodiments of this application can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code downloaded over a network that is originally stored in a remote recording medium or a non-transitory machine-readable medium and will be stored in a local recording medium. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0098] Those skilled in the art will recognize that the units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this application.

[0099] The above embodiments are only used to illustrate the embodiments of this application, and are not intended to limit the embodiments of this application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of this application. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of this application, and the patent protection scope of the embodiments of this application should be defined by the claims.

Claims

1. A non-destructive screening method for gene-edited resistant rice lines based on deep learning, characterized in that, The method is implemented using a rice disease resistance gene editing effect detection model. This model includes a ResNet50 encoder, a decoder with a morphological prior module and a multi-scale context-aware feature calibration module, a segmentation head, and a discrimination head. The method comprises: T0 generation regenerated plants obtained by CRISPR / Cas9 vector transformation and differentiation culture were obtained. The T0 generation regenerated plants were inoculated with bacterial blight pathogen. After inoculation, images of rice leaves to be tested were collected. The preprocessed images of rice leaves to be tested were input into the ResNet50 encoder to obtain five coding feature maps at different scales. The decoder's multi-scale context-aware feature calibration module dynamically calibrates the encoded feature maps at different scales and generates a fourth-level decoded feature map based on the calibrated features. The decoder's morphological prior module extracts features from the second-level decoded initial feature map to obtain a six-dimensional morphological feature vector. The second-level decoded initial feature map is generated based on the third-level decoded feature map and the calibrated features of the encoded feature map at the third scale. The last level of the four-level decoding feature map is input into the segmentation head to obtain a pixel-level disease segmentation probability map; the six-dimensional morphological feature vector and the bottleneck feature map are input into the identification head to obtain a resistance level probability distribution map, and the category corresponding to the highest probability in the resistance level probability distribution map is taken as the target resistance level; based on the target resistance level, it is determined whether the T0 generation regenerated plant carries a valid Xa21 gene editing event; wherein, the bottleneck feature map is the feature map with the smallest spatial resolution and the most channels among multiple coding feature maps of different scales; The six-dimensional morphological feature vector, pixel-level disease segmentation probability map, and target resistance level of positive lines identified as carrying valid editing events are visualized and displayed. The positive lines are marked as candidate lines to enter the T1 generation homozygous screening, and plants identified as negative or invalid editing are removed from the screening process.

2. The method according to claim 1, characterized in that, The CRISPR / Cas9 vector contains a gRNA expression cassette targeting the Xa21 gene and a Cas9 protein expression cassette. The target sequence of the gRNA expression cassette is located in the coding region or promoter region of the Xa21 gene, and is used to induce loss-of-function mutations or changes in promoter activity of the Xa21 gene.

3. The method according to claim 1, characterized in that, The ResNet50 encoder includes five processing stages, each of which outputs an encoded feature map at one scale; the decoder includes four multi-scale context-aware feature calibration modules. The multi-scale context-aware feature calibration module of the decoder dynamically calibrates the encoded feature maps at different scales and generates a four-level decoded feature map based on the calibrated features, including: The i-th multi-scale context-aware feature calibration module concatenates the input feature map of the i-th upsampling stage and the encoding feature map of the i-th scale along the channel dimension, and generates the i-th spatial attention feature through a 3x3 convolutional layer and a sigmoid function. Global average pooling is then applied to the input feature map of the i-th upsampling stage and the encoding feature map of the i-th scale, and after concatenation, the i-th channel attention feature is generated through a multilayer perceptron and a sigmoid function. The i-th spatial attention feature and the i-th channel attention feature are used to element-wise weight the encoding feature map of the i-th scale, and residual connections are performed to obtain the i-th calibrated feature. Based on the i-th calibrated feature and the input feature map of the i-th upsampling stage, the i-th level decoding feature map is generated. Where i is an integer greater than 0 and less than 5.

4. The method according to claim 2, characterized in that, The step of using the i-th spatial attention feature and the i-th channel attention feature to perform element-wise weighting of the encoded feature map at the i-th scale, and then performing residual connections to obtain the i-th calibrated feature, includes: The i-th spatial attention feature is reshaped and broadcast in the spatial dimension to obtain the i-th reference spatial attention feature map; Broadcast the attention feature of the i-th channel in the spatial dimension to obtain the attention feature map of the i-th reference channel; The i-th reference space attention feature map, the i-th reference channel attention feature map, and the i-th scale coding feature map are multiplied element-wise, and the i-th scale coding feature map is added as a residual term to obtain the i-th calibrated feature.

5. The method according to claim 2, characterized in that, The process of generating the initial feature map for the second-level decoding includes: The third-level decoded feature map is upsampled by a factor of two and processed by a 1x1 convolutional layer to obtain the guiding feature map; the number of channels in the guiding feature map is the same as the number of channels in the third-scale encoded feature map; The third multi-scale context-aware feature calibration module is used to dynamically calibrate the encoded feature map at the third scale to obtain the third calibrated feature. The guiding feature map and the third calibrated feature map are concatenated along the channel dimension and processed through two 3x3 convolutional layers to generate the initial feature map for the second-level decoding.

6. The method according to claim 1, characterized in that, The morphological prior module of the decoder extracts features from the initial feature map of the second-level decoding to obtain a six-dimensional morphological feature vector, including: The initial feature map of the second-level decoding is processed through a 1x1 convolutional layer using a morphological prior module to obtain a reference feature map. Dilation and erosion operations are then performed on the reference feature map to obtain dilated and eroded feature maps, respectively. A soft thresholding function is used to process the dilated and eroded feature maps to obtain dilated lesion probability maps and eroded lesion probability maps, respectively. Three morphological index values ​​are extracted from the dilated and eroded lesion probability maps using differentiable statistical methods. A six-dimensional morphological feature vector is constructed based on the extracted three morphological index values. Among them, the three morphological indicators include average lesion area, number of lesions, and lesion dispersion.

7. The method according to claim 6, characterized in that, The method further includes: The enhanced feature map is obtained by adding the reference feature map, dilation feature map, and erosion feature map element by element. The number of channels in the enhanced feature map is adjusted by using a 1x1 convolutional layer to obtain the input feature map for the second upsampling stage; Based on the input feature map from the second upsampling stage, the second multi-scale context-aware feature calibration module, and the second-scale encoded feature map, a first-level decoded feature map is generated.

8. The method according to claim 1, characterized in that, The last level of the four-level decoded feature map is input into the segmentation head to obtain a pixel-level disease segmentation probability map, including: The last level of the four-level decoded feature map is convolved using the 1x1 convolution module in the segmentation head to obtain a two-channel feature map. Then, the pixel-level disease segmentation probability map is generated by the Softmax function in the segmentation head.

9. The method according to claim 1, characterized in that, By inputting the six-dimensional morphological feature vector and the bottleneck feature map into the discrimination head, a probability distribution map of the resistance level is obtained, including: The bottleneck feature map is subjected to global average pooling using the global average pooling module in the discrimination head to obtain the global feature vector; The global feature vector and the six-dimensional morphological feature vector are concatenated by the feature concatenation module in the identification head to obtain a comprehensive feature vector; The comprehensive feature vector is processed using three fully connected layers and the Softmax function in the discrimination head to obtain the probability distribution map of resistance level.

10. The method according to claim 1, characterized in that, The training process of the rice disease resistance gene editing effect detection model includes: Images of rice leaf samples that have undergone rice disease resistance gene editing were collected, and pixel-level binary mask images and resistance level labels annotated by experts were obtained; Rice leaf sample images are input into the initial rice disease resistance gene editing effect detection model to obtain pixel-level disease segmentation prediction probability map and resistance level prediction probability distribution map; Based on the pixel-level disease segmentation prediction probability map, the resistance level corresponding to the maximum probability value in the resistance level prediction probability distribution map, the binary mask image, the resistance level label, and the total loss function constructed from segmentation loss and identification loss, the network parameters of the initial rice disease resistance gene editing effect detection model are backpropagated and trained until the convergence condition is met, thus obtaining the trained rice disease resistance gene editing effect detection model.

Images