Inplausible lifetime risk prediction method based on concept driving and visual language model

Abstract

The invention discloses an interpretable lifetime risk prediction method based on concept driving and a visual language model. The method comprises the following steps: 1, acquiring pathological concepts related to lifetime risk prediction; 2, designing positive and negative prompts in combination with a pathological concept, adopting a multi-prompt consensus strategy, and calculating a concept score of each image instance in the full-slice image by utilizing a visual language model; 3, the concept scores of the image instances are fused through a space perception Transform network, and slice-level concept scores are generated; and 4, inputting the slice-level concept score into a linear survival predictor for lifetime risk prediction. According to the method, traditional black box type deep learning prediction is converted into a transparent reasoning process based on pathological concepts, conceptualized interpretation consistent with clinical diagnosis logic can be provided while high prediction precision is guaranteed, the credibility and clinical acceptability of results are remarkably enhanced, and practical application of AI in pathological prognosis evaluation is promoted.

Description

Technical Field

[0001] This invention belongs to the fields of medicine and bioinformatics, specifically an interpretable survival risk prediction method based on concept-driven and visual language models. Background Technology

[0002] Survival risk prediction is of great significance in many fields, including medicine and biology. Current clinicopathological features have limitations in accurately predicting individual survival outcomes. Transcriptomics-based molecular typing methods are difficult to popularize in clinical practice due to their high cost and complex procedures. While deep learning-based whole-slice image analysis methods can predict survival risk, their models are often "black boxes," with opaque decision-making processes leading to low clinical trust. Although key areas can be highlighted using post-hoc interpretation methods such as attention mechanisms, these interpretations are disconnected from the professional concepts of pathologists and fail to provide truly clinically meaningful decision-making basis. In recent years, visual language models have demonstrated the ability to understand semantic concepts from pathological images, laying the foundation for building interpretable models. However, effectively integrating visual language models, which are only applicable to local image instances, into survival risk prediction for billion-pixel-level whole-slice images remains a pressing technical challenge. Summary of the Invention

[0003] This invention aims to address the shortcomings of existing technologies by proposing an interpretable survival risk prediction method based on concept-driven and visual language models. This method aims to improve the accuracy and interpretability of survival risk prediction, thereby solving problems such as opaque model decision-making and difficulty in gaining the trust of clinicians in existing technologies.

[0004] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides an interpretable survival risk prediction method based on concept-driven and visual language models, characterized by the following steps: Step 1, Obtain A pathological concept used for survival risk prediction, wherein any first... Each pathological concept is denoted as , ∈[1, ]; After obtaining a full-slice image W from the training dataset and performing tissue segmentation and partitioning, the preprocessed image is obtained. The nth image instance, where any nth Each image instance is denoted as , ∈[1, Let any number of... Image instances The spatial coordinates are ( , Let W's actual survival time be... ; Based on the true survival time distribution of all full-slice images where endpoint events occur in the training dataset, N+1 ordered time boundary points are determined. , , ,..., , ;in, This represents the (N-1)th time boundary point. Represents the initial time boundary point, and , Denotes the Nth time boundary point, and ; Step 2, based on Given a pathological concept, set K pairs of positive and negative prompt templates, where any ______... The positive prompt template in the positive and negative prompt templates is denoted as... The negative prompt template is recorded as , ∈[1,K]; Based on the A pathological concept , No. A positive reminder template and the Negative reminder template Using equations (1) and (2), we obtain the first... A pathological concept In the Positive descriptions in positive and negative prompt templates and negative descriptions : = ∪ (1) = ∪ (2) Step 3: Text encoder using a visual language model To each and Processing is performed, and the corresponding results are obtained. Text features and Text features ; Image encoder using visual language model right Processing is performed to obtain Image features ; Step 4: Using equations (3) and (4), obtain the first... Image instances Corresponding to the A pathological concept In the A positive prompt template Positive concept score and the Image instances Corresponding to the A pathological concept In the Negative reminder template Negative concept score : (3) (4) In equations (3) and (4), It is a cosine similarity function; Step 5: Filter using a multi-prompt consensus strategy and The low confidence score in the middle is the first Image instances In the A pathological concept High confidence score ; Step 6, according to Spatial coordinates ( , ) and the Image instances Spatial coordinates ( , ), thus obtaining the spatial prior attention matrix The Middle Line 1 Column elements ; ∈[1, ], ; Step 7, and The data are input into a spatially aware Transformer network for processing, and the corresponding results are obtained. Image instances In the A pathological concept Final score and slice-level concept scores ; Step 8, The input is processed by a linear survival predictor to obtain the final risk score of the whole-slice image W. ; Step 9, based on N+1 ordered time boundary points and Construct the total loss function and minimize To achieve this goal, the Adam optimizer is used to train an interpretable survival risk prediction network consisting of a spatially aware Transformer network and a linear survival predictor, thereby obtaining a trained interpretable survival risk prediction model for predicting the survival risk of an input full-slice image.

[0005] The characteristic of the interpretable survival risk prediction method based on concept-driven and visual language models described in this invention is that step 5 is performed as follows: Step 5.1: Use equation (5) to obtain the first... Image instances In the A pathological concept Based on the first Binarization prediction results of the prompt template = [ ] ,in, Indicates the first For the i-th image instance under the prompt template There exists a j-th pathological concept. The prediction results Indicates the first For the i-th image instance under the prompt template There is no j-th pathological concept. Prediction results: (5) In equation (5), I (·) is an indicator function that takes the value 1 when the condition is true and 0 otherwise. Step 5.2: Using equations (6) and (7), obtain the first... Image instances In the A pathological concept Binarized prediction probability vector on =[ ] and the Image instances In the A pathological concept Uncertainty measure ,in, Represents the i-th image instance There exists a j-th pathological concept. The probability, Represents the i-th image instance There is no j-th pathological concept. The probability of: (6) (7) Step 5.3: Use equation (8) to obtain the first... Image instances In the A pathological concept High confidence score Thus, the image instance concept score matrix is ​​obtained. : (8) In equation (8), This is the preset confidence threshold.

[0006] Furthermore, step 6 is performed as follows: Step 6.1: Using equation (9) to obtain and Manhattan distance : (9) Step 6.2, using equation (10) to obtain : (10) In equation (13), γ(·) is the negative correlation function.

[0007] Furthermore, step 7 is performed as follows: Step 7.1: Obtain the spatial context-enhanced image instance concept score matrix using equation (11). : (11) In equation (11), and For two linear mapping functions; For transpose; Indicates the activation function; Step 7.2: Obtain the first result using equation (12). Image instances Spatial perception attention weights : (12) In equation (12), , These are two parameters to be learned. express The first in row element, It is a cosine similarity function; Step 7.3: Obtain the slice-level concept score using equation (13). : (13).

[0008] Furthermore, in step 8, the final risk score is obtained using equation (14). (14) In equation (14), For the first A pathological concept The weight vector, For bias, It is a nonlinear activation function for risk normalization.

[0009] Furthermore, step 9 is performed as follows: Step 9.1: Based on N+1 ordered time boundary points, define N time intervals, where any nth interval is denoted as... , , ∈[1, ]; Step 9.2: Use equation (15) to obtain the discretized survival time label of the whole slice image W. : , (15) Step 9.3: Use equations (16) and (17) to obtain the full-slice image W in the first step. Conditional risk probability of the endpoint event occurring in each interval and survival time exceeding the first Survival probability in each interval : (16) (17) In equations (16) and (17), Represents a conditional probability function; Indicates the first The conditional risk probability of the endpoint event occurring in each interval; Step 9.4: Construct the total loss function using equation (18) : (18) In equation (18), Let W be the censored state corresponding to the full slice image W. When W is in the censored state, let... =1, when W is in an uncensored state, let =0; Indicates W in The conditional risk probability of the endpoint event occurring within the specified timeframe. This indicates that W's survival time exceeds The cumulative survival probability.

[0010] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in performing the method described therein, and the processor is configured to execute the program stored in the memory.

[0011] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to perform the steps of the method described thereon.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention proposes a concept-driven, interpretable survival risk framework: It innovatively integrates a visual language model with a multi-instance learning framework, using pathological concepts defined by pathology experts as a bridge, transforming deep learning models from a "black box" into a decision-making process that pathologists can understand and verify, significantly improving the interpretability and trustworthiness of clinical decisions.

[0013] 2. This invention achieves precise concept quantification from image instances to slices: By using positive and negative cue templates, multi-cue consensus strategies, and spatially aware Transformer networks, it robustly calculates the concept scores of local image instances and integrates spatial relationships to generate slice-level concept scores with clear pathological semantic associations, providing traceable pathological semantic features for survival prediction.

[0014] 3. This invention constructs an end-to-end interpretable prediction process: a linear survival predictor is used to map slice-level concept scores to risk scores, ensuring end-to-end transparency of the prediction process; each prediction can be traced back to a specific pathological concept and its spatial distribution, which can directly assist pathologists in predicting survival risk, greatly enhancing the clinical practical value and scientific research potential of the method. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0017] In this embodiment, an interpretable survival risk prediction method based on concept-driven and visual language models includes: pathological concept construction, image instance concept scoring, slice-level concept score generation, and linear survival prediction. Specifically, such as... Figure 1 As shown, it includes the following steps: Step 1, Obtain A pathological concept used for survival risk prediction, wherein any first... Each pathological concept is denoted as , ∈[1, ]; After acquiring a full-slice image W and performing tissue segmentation and division, the preprocessed image is obtained. The nth image instance, where any nth Each image instance is denoted as , ∈[1, Let any number of... Image instances The spatial coordinates are ( , Let W's actual survival time be... ; Based on the true survival time distribution of all full-slice images where endpoint events occur in the training dataset, N+1 ordered time boundary points are determined. , , ,..., , ;in, This represents the (N-1)th time boundary point. Represents the initial time boundary point, and , Denotes the Nth time boundary point, and .

[0018] Step 2, based on Given a pathological concept, set K pairs of positive and negative prompt templates, where any ______... The positive prompt template in the positive and negative prompt templates is denoted as... The negative prompt template is recorded as , ∈[1,K]; specifically, any ∈[1,K] A sample positive message template could be "An H&E image of {keyword}", and any second... A negative prompt template could be exemplified as "An H&E image of {no keyword}", where "keyword" can be replaced with any concept. Based on the... A pathological concept , No. A positive reminder template and the Negative reminder template Using equations (1) and (2), we obtain the first... A pathological concept In the Positive descriptions in positive and negative prompt templates and negative descriptions : = ∪ (1) = ∪ (2) Step 3: Text encoder using a visual language model To each and Processing is performed, and the corresponding results are obtained. Text features and Text features Specifically, , ; Image encoder using visual language model right Processing is performed to obtain Image features .

[0019] Step 4: Using equations (3) and (4), obtain the first... Image instances Corresponding to the A pathological concept In the A positive prompt template Positive concept score and the Image instances Corresponding to the A pathological concept In the Negative reminder template Negative concept score . and Used respectively for quantification of the first Image instances Corresponding to the A pathological concept In the The degree of conceptual representation under positive and negative prompt templates. When =1 This indicates that the target concept has the highest degree of presence in the image instance; when =1 This indicates that the target concept is most missing in this image instance; when and When the value approaches 0.5, it indicates that the visual language model cannot reliably determine whether the target concept exists in the image instance. (3) (4) In equations (3) and (4), It is a cosine similarity function.

[0020] Step 5: Filter using a multi-prompt consensus strategy and The low confidence score in the middle is the first Image instances In the A pathological concept High confidence score ; Step 5.1: Use equation (5) to obtain the first... Image instances In the A pathological concept Based on the first Binarization prediction results of the prompt template = [ ] ,in, Indicates the first Image instance under the prompt template The concept of pathology exists in China. The prediction results Indicates the first Image instance under the prompt template The concept of pathology does not exist in China. Prediction results: (5) In equation (5), I(·) This is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0021] Step 5.2: Using equations (6) and (7), obtain the first... Image instances In the A pathological concept Binarized prediction probability vector on =[ ] and the Image instances In the A pathological concept Uncertainty measure ,in, Represents an image instance The concept of pathology exists in China. The probability, Represents an image instance The concept of pathology does not exist in China. The probability of: (6) (7) Step 5.3: Use equation (8) to obtain the first... Image instances In the A pathological concept High confidence score Thus, the image instance concept score matrix is ​​obtained. Among them, when uncertainty is measured... satisfy When the concept score of the image instance is determined to be a high-confidence result, the average concept score of the K corresponding prompt templates for that image instance is taken to reduce random error and improve score stability; when the uncertainty measure satisfy When the concept score of an image instance is determined to be low confidence, the concept score of that image instance is uniformly set to 0.5. This corresponds to the ambiguous state where the visual language model cannot reliably determine the existence of the target concept, thus avoiding interference from low confidence data in subsequent analysis. (8) In equation (8), This is the preset confidence threshold.

[0022] Step 6, according to Spatial coordinates ( , ) and the Image instances Spatial coordinates ( , ), thus obtaining the spatial prior attention matrix The Middle Line 1 Column elements ; ∈[1, ], ; Step 6.1: Using equation (9) to obtain and Manhattan distance : (9) Step 6.2, using equation (10) to obtain : (10) In equation (13), γ(·) is the negative correlation function.

[0023] Step 7, and The data are input into a spatially aware Transformer network for processing, and the corresponding results are obtained. Image instances In the A pathological concept Final score and slice-level concept scores ; Step 7.1: Obtain the spatial context-enhanced image instance concept score matrix using equation (11). : (11) In equation (11), and For two linear mapping functions; For transpose; This represents the activation function.

[0024] Step 7.2: Obtain the first result using equation (12). Image instances Spatial perception attention weights : (12) In equation (12), , These are two parameters to be learned. Representation matrix The row element, It is a cosine similarity function; Step 7.3: Obtain the slice-level concept score using equation (13). : (13).

[0025] Step 8, The input is processed by a linear survival predictor to obtain the final risk score of the whole-slice image W. : Step 8.1, Input the linear survival predictor to obtain the final risk score. (14) In equation (14), For the first A pathological concept The weight vector, For bias, For risk normalization, a nonlinear activation function is used. It can be decomposed into the first Image instances In the A pathological concept High confidence score A linear combination of these elements maintains the model's complete transparency and interpretability.

[0026] Step 9, based on N+1 ordered time boundary points and Construct the total loss function and minimize To achieve this goal, the Adam optimizer is used to train an interpretable survival risk prediction network consisting of a spatially aware Transformer network and a linear survival predictor, thereby obtaining a trained interpretable survival risk prediction model for predicting the survival risk of an input full-slice image.

[0027] Step 9.1: Based on N+1 ordered time boundary points, define N time intervals, denoted as (the nth interval is not specified in the original text). , , ∈[1, ].

[0028] Step 9.2: Use equation (15) to obtain the discretized survival time label of the whole slice image W. : , (15) Step 9.3: Use equations (16) and (17) to obtain the full-slice image W in the first step. Conditional risk probability of the endpoint event occurring in each interval and survival time exceeding the first Survival probability in each interval : (16) (17) In equations (16) and (17), This represents a conditional probability function; specifically... This indicates that the known survival time is not less than the [number]th [number]. Given a given interval and a final risk score, at the [number]th interval, [the following is a possible interpretation of the given information]. The probability of the endpoint event occurring within a given interval. Indicates the first The probability of a conditional risk of the endpoint event occurring in each interval.

[0029] Step 9.4: Construct the total loss function using equation (18) : (18) In equation (18), For the censored state corresponding to the full slice image W, when the data is censored =1, when the data is not censored. =0. Indicates that W is in the interval The conditional risk probability of the endpoint event occurring within the specified timeframe. This indicates that W's survival time exceeds the interval. The cumulative survival probability.

[0030] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the methods described above, and the processor is configured to execute the program stored in the memory.

[0031] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0032] To quantitatively evaluate the performance of the proposed method, this embodiment demonstrates its comparison with concept-based methods LaBo (Yang, Yue, et al. "Language in a bottle: Language model guided concept bottlenecks for interpretable image classification." Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2023.), MONET (Kim, Chanwoo, et al. "Transparent medical image AI via an image–text foundation model grounded in medical literature." Nature medicine 30.4 (2024): 1154-1165.) and multi-instance learning methods TransMIL (Shao, Zhuchen, et al. "Transmil: Transformer-based correlated multiple instance learning for whole slide image classification." Advances in neural information processing systems 34(2021): 2136-2147.), HIPT (Chen, Richard J., et al. "Scaling vision transformers to gigapixel images via hierarchical self-supervised learning."Proceedings of the IEEE / CVF conference on computer vision and patternrecognition. 2022.), ACMIL (Zhang, Yunlong, et al. "Attention-challengingmultiple instance learning for whole slide image classification." Europeanconference on computer vision.The performance comparison results of (Cham: Springer Nature Switzerland, 2024.) on the internal dataset Internal-CMU1H and the public dataset External-TCGA are shown in Table 1.

[0033] Table 1: Performance comparison of the proposed method with concept-based methods and multi-instance learning methods on the internal dataset Internal-CMU1H and the public dataset External-TCGA.

[0034] Table 1 presents a performance comparison between the proposed method CDM and existing baseline models in the survival risk prediction task, and provides a systematic overview of the experimental results. The experimental results show that, within the internal CMU1H cohort, CDM achieves a C-index of 0.779 (95% CI: 0.768–0.789) and an AUC of 0.813 (95% CI: 0.791–0.834). Notably, CDM significantly outperforms traditional concept-based models (LaBo: C-index 0.691; MONET: C-index 0.698), with relative improvements of +8.8% and +8.1%, respectively. Its performance is similar to the best-performing black-box MIL method ACMIL (C-index: 0.782; AUC: 0.818). A consistent trend was observed in the external TCGA validation queue, where CDM achieved a C-index of 0.741 (95% CI: 0.698–0.782) and an AUC of 0.800 (95% CI: 0.754–0.842) (Figure S1), demonstrating excellent generalization performance and competitive performance with ACMIL (C-index: 0.758; AUC: 0.821). In summary, CDM achieves a good balance between performance and interpretability.

Claims

1. An interpretable survival risk prediction method based on concept-driven and visual language model, characterized in that, is performed by the following steps: Step 1, obtaining a pathology concept for survival risk prediction, wherein any i-th pathology concept is denoted by , ∈ [1, ] ; After obtaining a whole slice image W in the training data set and performing tissue segmentation and cutting, a preprocessed image instance is obtained, wherein any i-th image instance is denoted as Wi, i ∈ [1, N], and the spatial position coordinates of any i-th image instance Wi are denoted as (xi, yi). , ; and a real survival time of W is denoted as T.​​​​​​​​ determining N+1 ordered time boundary points based on a true survival time distribution of all full slice images in the training data set that experienced an end point event , , , , ; wherein denotes the (N-1)-th time boundary point, denotes the initial time boundary point, and , denotes the N-th time boundary point, and ; Step 2, based on K pairs of positive and negative prompt templates are set according to the pathological concept, wherein any i-th positive prompt template in the positive and negative prompt templates is denoted as , and the negative prompt template is denoted as , ∈[1,K] Based on the A pathological concept , No. Positive Reminder Template and the Negative reminder template Using equations (1) and (2), we obtain the first... A pathological concept In the Positive descriptions in positive and negative prompt templates and negative descriptions : = ∪ (1) = ∪ (2) Step 3: Text encoder using a visual language model To each and Processing is performed, and the corresponding results are obtained. Text features and Text features ; Image encoder utilizing visual language model processing, obtaining image features of ;​​ Step 4: Using equations (3) and (4), obtain the first... Image instances Corresponding to the A pathological concept In the A positive prompt template Positive concept score and the Image instances Corresponding to the A pathological concept In the Negative reminder template Negative concept score : (3) (4) In formula (3) and formula (4), is a cosine similarity function; Step 5, filtering with multi-prompt consensus strategy and low confidence scores to get the image instance with high confidence score of the pathological concept ;​ Step 6, according to Spatial coordinates ( , ) and the Image instances Spatial coordinates ( , ), thus obtaining the spatial prior attention matrix The Middle Line 1 Column elements ; ∈[1, ], ; Step 7, inputting the first image instance into the spatial-aware Transformer network for processing, to obtain a first image instance feature vector and respectively input into the spatial-aware Transformer network for processing, to obtain a corresponding first image instance feature vector and a second image instance feature vector respectively a final score of the first pathological concept and a slice-level concept score and a slice-level concept score ; Step 8, the The input linear survival predictor is processed to obtain the final risk score of the whole slice image W ; Step 9, based on N+1 ordered time boundary points and constructing a total loss function and taking minimizing as the goal, using an Adam optimizer to train the interpretable survival risk prediction network composed of the spatial perception Transformer network and the linear survival predictor, so as to obtain a trained interpretable survival risk prediction model for performing survival risk prediction on the input whole slide image.

2. The concept-driven and visual language model-based interpretable survival risk prediction method according to claim 1, characterized in that, The step 5 is performed by the following steps: Step 5.1: Use equation (5) to obtain the first... Image instances In the A pathological concept Based on the first Binarization prediction results of the prompt template = [ ] ,in, Indicates the first For the i-th image instance under the prompt template There exists a j-th pathological concept. The prediction results Indicates the first For the i-th image instance under the prompt template There is no j-th pathological concept. Prediction results: (5) In formula (5), I (·) is an indicator function that takes the value 1 when the condition is true and 0 otherwise. Step 5.2: Using equations (6) and (7), obtain the first... Image instances In the A pathological concept Binarized prediction probability vector on =[ ] and the Image instances In the A pathological concept Uncertainty measure ,in, Represents the i-th image instance There exists a j-th pathological concept. The probability, Represents the i-th image instance There is no j-th pathological concept. The probability of: (6) (7) Step 5.3, obtaining the first image instance using formula (8) In the second pathological concept A high confidence score is obtained from the first image instance concept , thereby obtaining a matrix of image instance concept scores :​​ (8) In formula (8), is a preset confidence threshold.

3. The concept-driven and visual language model-based interpretable survival risk prediction method of claim 1, wherein, The step 6 is performed by the following steps: Step 6.1, obtaining a Manhattan distance between and :​ (9) Step 6.2, using formula (10) to obtain : (10) In formula (13), γ(·) is a negative correlation function.

4. The concept-driven and visual language model-based interpretable survival risk prediction method of claim 2, wherein, The step 7 is performed by the following steps: Step 7.1, obtaining a spatially context enhanced image instance concept score matrix using formula (11) : (11) In formula (11), and are two linear mapping functions; is a transpose; denotes an activation function; Step 7.2, obtaining the first image instance by formula (12) :​​ (12) In formula (12), , are two parameters to be learned, denotes the element in the row, is a cosine similarity function; Step 7.

3. Obtain slice level concept score by formula (13) : (13)。 5. The concept-driven and visual language model-based interpretable survival risk prediction method of claim 1, wherein, The step 8 is to obtain the final risk score using formula (14) (14) In formula (14), is the weight vector for the pathological concept , is the bias, is the risk-normalized nonlinear activation function.

6. The concept-driven and visual language model-based interpretable survival risk prediction method of claim 1, wherein, The step 9 is performed by the following steps: Step 9.

1. Define N time intervals based on N+1 ordered time boundary points, where any nth interval is denoted as , , ∈ [1, ]. Step 9.2, obtaining a discretized lifetime label for the full slice image W using formula (15) : ， (15) Step 9.

3. Obtain the full slice image W at the end of the interval using (16) and (17) the conditional risk probability that the end of life event occurs in the interval and the survival probability that the survival time exceeds the interval :​ (16) (17) In formula (16) and formula (17), denotes a conditional probability function; denotes a conditional risk probability of an end event occurring in the th interval. Step 9.

4. Constructing the overall loss function with formula (18) : (18) In formula (18), is the censoring status corresponding to the full slice image W, when W is a censored status, let = 1, when W is an uncensored status, let = 0; represents the conditional risk probability of W having an endpoint event within , represents the cumulative survival probability of the survival time of W exceeding .

7. An electronic device comprising a memory and a processor, characterized in that The memory is configured to store a program supporting the processor to execute the method of any one of claims 1-6.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program, when executed by a processor, performs the steps of the method of any one of claims 1-6.

Images