A non-invasive ATP prediction method for organoids based on deep learning
By using a U-shaped neural network structure based on deep learning, non-invasive, real-time, and high-throughput prediction of organoid ATP can be achieved based on bright-field microscopic images. This solves the problems of invasiveness and low throughput of traditional methods and is suitable for organoid viability assessment and biomedical research.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUDAN UNIVERSITY
- Filing Date
- 2026-01-20
- Publication Date
- 2026-05-26
Smart Images
Figure CN122090127A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of organoid image analysis technology, specifically relating to an organoid ATP prediction method based on bright-field microscopy images. Background Technology
[0002] Organoids, as three-dimensional cell cultures, can simulate the structure and function of real organs, and have important applications in drug screening and disease modeling. In recent years, organoids have shown great potential in predicting drug responses and personalized treatment for cancer, providing a new research platform for precision medicine by simulating the genetic, phenotypic, and microenvironmental characteristics of patients.
[0003] In organoid-based drug screening and disease modeling, organoid viability assessment is a crucial step. Adenosine triphosphate (ATP), as the "gold standard" indicator of cellular energy metabolism, is widely used for assessing organoid viability. [1] However, traditional ATP detection methods are invasive, requiring the use of biochemical reagents to disrupt cell structure. This limits their application to endpoint detection and makes it difficult to achieve dynamic and continuous monitoring of organoid ATP. Furthermore, these methods suffer from low throughput, hindering their ability to support long-term, multi-sample ATP monitoring tasks.
[0004] Organoid image analysis is a non-invasive method for assessing organoid viability. With the development of live-cell imaging and quantitative microscopy, researchers can obtain morphological and textural features of organoids through bright-field microscopy images. These features are related to the physiological state of organoids and can be used to infer functional indicators such as cell proliferation, apoptosis, and metabolic activity. In recent years, deep neural network methods have made significant progress in the application of medical image analysis, especially in regression tasks, demonstrating the potential to predict continuous biological variables. Nevertheless, there is still a lack of methods that can effectively predict organoid ATP. Existing methods are mostly focused on classification tasks or applied to areas that differ significantly from organoid image features, making it difficult to directly generalize to organoid ATP prediction tasks. Summary of the Invention
[0005] The purpose of this invention is to propose an efficient, non-destructive, and dynamic deep learning-based non-invasive method for predicting organoid ATP.
[0006] This invention provides a non-invasive ATP prediction method for organoids based on deep learning, comprising: preprocessing bright-field microscopic images, such as normalization, to obtain input images that can be directly used by the model; employing a U-shaped neural network structure suitable for regression tasks, the encoder of which consists of three parallel convolutional branches, which extract local texture features and global contextual features from the image respectively, and reduce feature redundancy; each branch ends with a channel attention weighting module to enhance the expressive power of key information. Through attention-enhanced skip connections, multi-scale features are passed to the decoder, and combined with a lightweight spatial compression module and a fully connected regression head, end-to-end prediction from microscopic images to ATP values is achieved. The model is trained under supervision based on organoid images labeled with ATP values, ultimately yielding a neural network model that can be used for lossless prediction of organoid ATP levels. This invention includes two core steps: image preprocessing, and the construction of an ATP prediction model based on a deep neural network for multi-branch feature extraction.
[0007] (a) Image preprocessing: The original brightfield micrograph is subjected to grayscale normalization and size uniformization to obtain an input image with consistent size and pixel distribution;
[0008] (II) Construction of the ATP prediction model: The ATP prediction model includes a multi-branch feature extraction module, a feature restoration module, and a bottleneck spatial compression regression module; wherein:
[0009] (1) Multi-branch feature extraction module (encoder) obtains multi-scale local texture, context and global background representation by using three parallel feature extraction paths, and introduces channel attention weighting after feature fusion to highlight channels related to ATP;
[0010] (2) Feature restoration module (decoder): While upsampling step by step, attention-gated skip connections are used to selectively inject the encoded features of the corresponding scale into the decoding path in order to preserve spatial details and suppress redundancy.
[0011] (3) Bottleneck spatial compression regression module, which maps high-dimensional features into a single ATP value by using bottleneck-style stepwise convolution compression and fully connected regression head;
[0012] The network uses real ATP values as a supervision signal and employs mean absolute error to optimize the model parameters.
[0013] The image obtained in step (1) is input into the ATP prediction model to achieve feature extraction, information fusion and numerical prediction.
[0014] Furthermore:
[0015] The image preprocessing described in step (I) specifically includes:
[0016] The input is uniformly scaled to a fixed resolution and then compressed in one go using 2×2 max pooling to form a standard input.
[0017] Subsequently, grayscale normalization and local contrast enhancement were performed to compensate for exposure differences during imaging;
[0018] Lightweight random rotation, flipping, and brightness perturbation are used to augment the samples, generating more training samples to improve the model's generalization ability.
[0019] In step (two):
[0020] The multi-branch feature extraction module (encoder) employs three parallel feature extraction paths to obtain multi-scale information, specifically including:
[0021] (1) First path: After two convolutional layers, perform a 2x downsampling to preserve medium-scale texture information;
[0022] (2) Second path: First perform 2x downsampling, then perform convolution processing to extract context structure information;
[0023] (3) Third path: First perform downsampling of no less than 4 times, then perform convolution processing, and align the output size through bilinear interpolation to capture the global structure and reduce noise interference.
[0024] The outputs of the three paths are concatenated along the channel dimension and then fused and compressed using a 3×3 convolution layer. The resulting data is then fed into a channel attention weighting module, which generates channel weights based on a global statistical method using average pooling and max pooling, automatically enhancing responses to important features relevant to ATP prediction. To avoid overfitting, a random neuron dropout mechanism is added after each path, with a dropout rate not exceeding 50%.
[0025] The feature restoration module (decoder) recovers image spatial information through a step-by-step upsampling operation, and its specific structure is as follows:
[0026] (1) The decoder adopts a three-level structure, each level including: a transposed convolution layer for resolution improvement, a normalization and activation function layer for introducing non-linear expression, and a convolution layer for refining feature maps;
[0027] (2) After each upsampling, the result is concatenated with the feature map of the corresponding scale in the encoder to form a skip connection structure, and important features are weighted again through the channel attention weighting module;
[0028] (3) The final output feature map is consistent with the input in terms of spatial size, and integrates deep semantic and shallow texture information to provide a complete expression for subsequent prediction.
[0029] Furthermore, the channel attention weighting module adopts a lightweight structure of double statistical concatenation and two-stage 1×1 convolution:
[0030] (1) Concatenation pooling unit: Perform global average pooling and max pooling on the input feature map, and concatenate in the channel direction to generate a feature map with twice the number of channels as the original;
[0031] (2) Convolution processing unit: Two 1×1 convolution operations are executed in sequence. The first layer compresses the number of channels to 1 / 4 of the original number of channels, and the second layer restores the original number of channels.
[0032] (3) Activation function processing: ReLU and Sigmoid activation functions are applied sequentially after the two convolution operations to generate the weight coefficients for each channel;
[0033] (4) Feature weighting operation: Multiply the weights element by element with the original features to complete the channel-level recalibration.
[0034] The bottleneck-type spatial compression regression module is used to map the feature map to the corresponding ATP value of the image, and its structure includes:
[0035] (1) Five convolutional units with a stride of 2, which compress the spatial dimension and increase the number of channels to 128 and then reduce it to 32;
[0036] (2) Flattening operation: convert the multidimensional feature map into a one-dimensional vector;
[0037] (3) A three-layer fully connected network with output dimensions of 512, 128 and 1 respectively. The first two layers use the ReLU activation function and finally output the ATP prediction value corresponding to the image.
[0038] The entire model is trained under supervision using actual ATP values as labels and optimized using mean absolute error (MAE) as the loss function. After training, it can predict organoid ATP values based on bright-field microscopic images.
[0039] Compared with existing technologies, this invention has the following advantages: it can directly predict the ATP level of organoids through bright-field microscopy images without the need for staining, lysis, or other destructive treatments of the samples, achieving non-invasive, real-time, and continuous viability assessment; based on a deep neural network architecture, the model can automatically extract multi-scale features, making it suitable for processing large batches of organoid samples; at the same time, this invention has the characteristics of high throughput and automation, making it suitable for biomedical research and preclinical evaluation scenarios such as drug screening and disease modeling, meeting the current development needs of precision medicine and personalized treatment. Attached Figure Description
[0040] Figure 1 This invention presents the ATP prediction model structure based on a multi-branch feature extraction deep neural network. Detailed Implementation
[0041] The following example, using bright-field microscopic images of lung cancer organoids obtained with the BioTek Cytation 7 imaging system, illustrates the practical application process of this invention. The resolution of the acquired images was 1992×1992, with a corresponding physical size of 3.45mm×3.45mm. After image acquisition, the corresponding ATP values were measured as the true labels.
[0042] This invention comprises two steps:
[0043] Step 1: Image Preprocessing
[0044] The original bright-field micrographs were first uniformly scaled to 1536×1536 pixels and then compressed to 768×768 pixels using a single 2×2 max pooling operation, which was then used as the standard input for the neural network model. To improve image stability and contrast, grayscale normalization and local histogram equalization were further performed on the images to eliminate exposure differences and enhance detail in dark areas.
[0045] During the training phase, to improve the model's robustness to multivariate imaging conditions, image augmentation is performed on the training samples, including random rotation, flipping, and slight brightness perturbations, in order to generate more samples and improve the model's generalization ability.
[0046] Step 2: Construction of ATP Prediction Model
[0047] The preprocessed image is input into the deep neural network model constructed in this invention. The overall structure of the model includes three main modules: a multi-branch feature extraction module, a feature restoration module, and a spatial compression regression module, as described below:
[0048] (a) Multi-branch feature extraction module (encoder)
[0049] This module consists of three parallel feature extraction paths, which extract image information from different scales, as shown in the following structure:
[0050] (1) The first path performs two 3×3 convolution operations before image downsampling to preserve high-resolution texture details;
[0051] (2) The second path first performs a 2x spatial downsampling, and then performs two convolution operations to extract medium-scale and contextual structure information;
[0052] (3) The third path directly performs 4 times spatial downsampling, then performs two convolution processes, and finally restores the features to the original scale through bilinear interpolation, which is used to capture the global structural features of the image and suppress noise interference.
[0053] The outputs of the three paths are concatenated along the channel dimension and then fused and compressed using a 3×3 convolution layer. The fused features are fed into a channel attention module, which generates statistical information using a concatenation of global average pooling and max pooling. Channel weights are generated through two 1×1 convolution layers and a sigmoid activation function, thereby enhancing the feature responses most relevant to ATP prediction. To prevent overfitting, a random neuron deactivation mechanism is added at the end of each path, with the proportion of dropped neurons not exceeding 50%.
[0054] (ii) Feature Reconstruction Module (Decoder)
[0055] This module is used to restore spatial resolution step by step, and its specific structure is as follows:
[0056] (1) The decoder contains three upsampling stages, each stage including a 2×2 transposed convolution for resolution amplification, a set of batch normalization and activation functions for stable training, and a 3×3 convolution to further refine the feature representation;
[0057] (2) The upsampled output of each level will be concatenated with the feature map of the corresponding scale in the encoder to construct a skip connection; the concatenation result will be further enhanced by the channel attention module to strengthen the effective information expression.
[0058] (3) The final output high-dimensional feature map is consistent with the input image in terms of spatial size, and integrates deep semantic and shallow texture features.
[0059] (III) Spatial Compression Regression Module
[0060] The feature map output from the decoder is input into the regression prediction module to estimate the ATP value. The process is as follows:
[0061] (1) The feature map first undergoes five levels of 3×3 convolution operations (with a stride of 2) for spatial compression. During the process, the number of channels gradually increases to 128 and then decreases to 32.
[0062] (2) The convolution output is flattened into a one-dimensional vector;
[0063] (3) The vector passes through two fully connected layers (dimensions of 512 and 128, activation function of ReLU) and one output layer in sequence to finally generate the ATP prediction value of the corresponding image.
[0064] The model uses mean absolute error (MAE) as the loss function and employs the Adam optimizer (with an initial learning rate of 5 × 10⁻⁶). -4The model was trained iteratively for 200 rounds on a training set, validation set, and test set with an 8:1:1 ratio. After training, the model was evaluated on 50 independent test samples, and the results are shown in Table 1. The mean absolute error (MAE) was 811.49, the root mean square error (RMSE) was 1030.86, and the coefficient of determination (R²) was 0.9872, demonstrating better performance compared to other comparative methods.
[0065] Table 1 shows the ATP prediction results obtained on 50 independent test set data.
[0066] .
[0067] The results show that the present invention can achieve rapid and accurate estimation of organoid ATP values under the condition of no ATP, and has broad application prospects. It is suitable for specific applications such as high-throughput drug screening and long-term dynamic monitoring of organoid function.
[0068] References
[0069] [1] Crouch SPM, Kozlowski R, Slater KJ, et al. The use of ATPbioluminescence as a measure of cell proliferation and cytotoxicity[J]. Journal of Immunological Methods, 1993, 160(1): 81-88.
[0070] [2] Bian X, Chen S, Liu W. Ins-ATP: Deep estimation of ATP fororganoid based on high throughput microscope images[J]. Methods, 2025, 235:34-44.
[0071] [3] He K, Zhang X, Ren S, et al. Deep residual learning for imagerecognition[C] / / Proceedings of the IEEE Conference on Computer Vision andPattern Recognition, 2016: 770-778.
[0072] [4] Ronneberger O, Fischer P, Brox T. U-net: Convolutional networksfor biomedical image segmentation[C] / / Proceedings of 2015 Conference ofMedical Image Computing and Computer-Assisted Intervention, 2015: 234-241。
Claims
1. A non-invasive method for predicting organoid ATP based on deep learning, characterized in that, include: Preprocessing is performed on the bright-field micrographs to obtain input images that can be directly used by the model; A U-shaped neural network structure suitable for regression tasks is adopted. Its encoder consists of three parallel convolutional branches, which extract local texture features and global context features from the image respectively, and reduce feature redundancy. Each branch ends with a channel attention weighting module to enhance the ability to express key information. Through attention-enhanced skip connections, multi-scale features are passed to the decoder. Combined with a lightweight spatial compression module and a fully connected regression head, end-to-end prediction from microscopic images to ATP values is achieved. The model is trained under supervision based on organoid images labeled with ATP values, and finally a neural network model is obtained for non-destructive prediction of organoid ATP levels.
2. The organoid ATP non-invasive prediction method according to claim 1, characterized in that, It includes two steps: (a) Image preprocessing: The original brightfield micrograph is subjected to grayscale normalization and size uniformization to obtain an input image with consistent size and pixel distribution; (II) Construction of an ATP prediction model based on a multi-branch feature extraction deep neural network, wherein the ATP prediction model includes a multi-branch feature extraction module, a feature restoration module, and a bottleneck spatial compression regression module; wherein: (1) Multi-branch feature extraction module: It obtains multi-scale local texture, context and global background representation by using three parallel feature extraction paths, and introduces a channel attention weighting module after feature fusion to highlight the channels related to ATP; (2) Feature restoration module: while upsampling step by step, attention-gated skip connections are used to selectively inject the encoded features of the corresponding scale into the decoding path in order to preserve spatial details and suppress redundancy. (3) Bottleneck spatial compression regression module, which maps high-dimensional features into a single ATP value by using bottleneck-style stepwise convolution compression and fully connected regression head; The network uses real ATP values as a supervision signal and optimizes the model parameters using mean absolute error. The image obtained in step (1) is input into the ATP prediction model to achieve feature extraction, information fusion and numerical prediction.
3. The organoid ATP non-invasive prediction method according to claim 2, characterized in that, The image preprocessing described in step (I) specifically includes: The input is uniformly scaled to a fixed resolution and then compressed in one go using 2×2 max pooling to form a standard input. Subsequently, grayscale normalization and local contrast enhancement were performed to compensate for exposure differences during imaging; Lightweight random rotation, flipping, and brightness perturbation are used to augment the samples, generating more training samples to improve the model's generalization ability.
4. The organoid ATP non-invasive prediction method according to claim 3, characterized in that, In step (ii), the multi-branch feature extraction module employs three parallel feature extraction paths to obtain multi-scale information, specifically including: (1) First path: After two convolutional layers, perform a 2x downsampling to preserve medium-scale texture information; (2) Second path: First perform 2x downsampling, then perform convolution processing to extract context structure information; (3) Third path: First perform downsampling of no less than 4 times, then perform convolution processing, and align the output size through bilinear interpolation to capture the global structure and reduce noise interference; The outputs of the three paths are concatenated along the channel dimension and then fused and compressed through a 3×3 convolution layer. The result is then fed into the channel attention weighting module, which generates channel weights based on a global statistical method that combines average pooling and max pooling. This automatically enhances the response to important features related to ATP prediction. To avoid overfitting, a random neuron drop-off mechanism is added after each path, with a drop-off rate not exceeding 50%.
5. The organoid ATP non-invasive prediction method according to claim 4, characterized in that, In step (ii), the feature restoration module recovers the image spatial information through a step-by-step upsampling operation. The specific structure is as follows: (1) The decoder adopts a three-level structure, each level including: a transposed convolution layer for resolution improvement, a normalization and activation function layer for introducing non-linear expression, and a convolution layer for refining feature maps; (2) After each upsampling, the result is concatenated with the feature map of the corresponding scale in the encoder to form a skip connection structure, and important features are weighted again through the channel attention weighting module; (3) The final output feature map is consistent with the input in terms of spatial size, and integrates deep semantic and shallow texture information to provide a complete expression for subsequent prediction.
6. The organoid ATP non-invasive prediction method according to claim 5, characterized in that, The channel attention weighting module adopts a lightweight structure of double statistical concatenation and two-stage 1×1 convolution: (1) Concatenation pooling unit: Perform global average pooling and max pooling on the input feature map, and concatenate in the channel direction to generate a feature map with twice the number of channels as the original; (2) Convolution processing unit: Two 1×1 convolution operations are executed in sequence. The first layer compresses the number of channels to 1 / 4 of the original number of channels, and the second layer restores the original number of channels. (3) Activation function processing: ReLU and Sigmoid activation functions are applied sequentially after the two convolution operations to generate the weight coefficients for each channel; (4) Feature weighting operation: Multiply the weights element by element with the original features to complete the channel-level recalibration.
7. The organoid ATP non-invasive prediction method according to claim 6, characterized in that, The bottleneck-type spatial compression regression module mentioned in step (II) specifically includes: (1) Five convolutional units with a stride of 2, which compress the spatial dimension and increase the number of channels to 128 and then reduce it to 32; (2) Flattening operation: convert the multidimensional feature map into a one-dimensional vector; (3) A three-layer fully connected network with output dimensions of 512, 128 and 1 respectively. The first two layers use the ReLU activation function and finally output the ATP prediction value corresponding to the image.
8. The organoid ATP non-invasive prediction method according to claim 7, characterized in that, The ATP prediction model is trained under supervision using actual ATP values as labels and optimized using mean absolute error (MAE) as the loss function. After training, organoid ATP value prediction based on bright-field microscopic images is achieved.