Method, system and equipment for evaluating language development of premature infant and medium

Abstract

The invention provides a premature infant language development evaluation method, system, equipment and medium, and belongs to the field of applying deep learning to neonatal brain function development disorder screening, and the method comprises the following steps: obtaining a brain image of a neonatal, extracting feature maps of the back curvature, the cortex grey matter thickness, the back depth and the myelin sheath, and obtaining a brain image of the neonatal; dividing the brain image into a plurality of triangular image blocks, and taking the feature maps of the back curvature, the cortex grey matter thickness, the back depth and the myelin sheath as feature channels of each vertex of each triangular image block to obtain feature embedding vectors of the left half brain and the right half brain; inputting the feature embedding vectors of the left half brain and the right half brain into a bidirectional scanning network for forward processing and reverse processing to obtain nonlinear features between the feature embedding vectors of the left half brain and the right half brain, and compressing the nonlinear features to a state space with a fixed size by using a state space model and outputting the nonlinear features. According to the invention, the language development evaluation efficiency of the premature infant can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of applying deep learning to the surface of the brain sphere, and specifically relates to a method, system, device and medium for evaluating the language development of premature infants. Background Art

[0002] Complete language skills are the basis for humans to engage in social activities. The age of 0 to 3 is a stage in which human language function develops rapidly. During this period, the human brain will form top-down language processing capabilities. The impact of premature birth on brain development, especially the damage to language function, has become an important research topic in clinical and scientific research fields in recent years. Compared with children born at full term, the brain development of premature infants is often seriously affected due to incomplete early development. At present, there is a lack of effective screening technology for the degree of language development of premature infants in clinical practice, making early identification and intervention particularly urgent.

[0003] However, there is still no mature model that can accurately predict and provide early diagnostic recommendations for predicting long-term language development impairment in premature infants. Although traditional statistical analysis methods can identify some risk factors, they are limited in predictive accuracy and efficiency when dealing with complex brain structural changes. With the continuous development of deep learning technology, researchers have begun to explore combining traditional data processing methods with new algorithms to improve the ability to analyze complex data. In order to solve these problems, some researchers proposed Surface Vision Transformer (SiT), which provides a new perspective for the feasible application of attention mechanism on irregular geometry. However, this method also has some defects, as follows:

[0004] First, SiT is computationally inefficient. Although it performs well in learning long-range dependencies, its self-attention mechanism requires pairing each input element with all other input elements. As the input size increases, the amount of computation increases quadratically, which increases the computational cost and leads to high computational complexity. It may be inefficient when processing large-scale data sets.

[0005] Secondly, SiT has high requirements for video memory. Although SiT can effectively learn important features in some cases, due to the existence of KVCache in the attention mechanism, the key-value pair of each feature will be retained, resulting in a large amount of video memory being consumed when processing long sequences.

[0006] In summary, the methods used in the prior art to predict the long-term language development of premature infants have problems of high computational complexity, high storage complexity and low computational efficiency, which affects the efficiency of language development assessment of premature infants. Summary of the invention

[0007] In order to overcome the above-mentioned shortcomings of low efficiency of existing premature infant language development assessment, the present invention provides a premature infant language development assessment method, comprising the following steps:

[0008] Obtain a newborn's brain image, extract the feature map of sulcus curvature, cortical gray matter thickness, sulcus depth and myelin sheath in the brain image, divide the brain image into multiple triangular image blocks, and use the feature map of sulcus curvature, cortical gray matter thickness, sulcus depth and myelin sheath as the feature channel of each vertex of each triangular image block to obtain the feature embedding vectors of the left and right hemispheres;

[0009] The feature embedding vectors of the left and right brain are input into a bidirectional scanning network, and the input data is forward processed and reverse processed to obtain the nonlinear features between the feature embedding vectors of the left and right brain. The state space model is used to compress the nonlinear features into a state space of a fixed size and output them. The language development of premature infants is evaluated based on the output data.

[0010] Preferably, the extraction of the characteristic map of sulcus curvature, cortical gray matter thickness, sulcus depth and myelin sheath in the brain image specifically comprises the following steps:

[0011] The Draw-EM algorithm was used to segment brain images into cerebrospinal fluid, white matter, cortical gray matter, and subcortical gray matter;

[0012] The pial surface was generated by extending the white matter mesh toward the cortical gray matter and cerebrospinal fluid boundaries, and the sulcal and gyral curvature was estimated from the white matter surface;

[0013] Cortical gray matter thickness was estimated based on the Euclidean distance between the white matter surface and the pial surface;

[0014] Dilated surfaces were generated by performing extension-based smoothing of the white matter surface, and sulcus-gyrus depths were estimated during dilation;

[0015] The ratio of weighted T1w and weighted T2w images was used to generate a characteristic map of myelin sheaths using FreeSurfer software.

[0016] Preferably, obtaining the feature embedding vectors of the left brain and the right brain comprises the following steps:

[0017] The left and right brain surfaces of the brain image were resampled onto a sixth-order regular icosahedron, and the left and right brain surfaces were divided into 1280 triangular image blocks, each containing 45 vertices.

[0018] The feature maps of sulcus curvature, cortical gray matter thickness, sulcus depth and myelin sheath are used as feature channels of each vertex of each triangular image block to obtain the feature representation of the left and right hemispheres.

[0019] The feature representations of the left and right hemispheres are flattened and input into the MLP to map them into feature embedding vectors of dimension D.

[0020] Preferably, before inputting the feature embedding vectors of the left brain and the right brain into the bidirectional scanning network, the method also includes splicing a learnable classifier between the feature embedding vector of the left brain and the feature embedding vector of the right brain and adding a one-dimensional position encoding.

[0021] Preferably, the bidirectional scanning network comprises 24 layers, each layer comprising two MLPs, two one-dimensional convolutions after one of the MLPs, two SiLU activation functions and two state space models and a SiLU activation function after another MLP.

[0022] Preferably, the forward processing and reverse processing of the input data are specifically: using one-dimensional convolution in each direction to establish local contextual relationships between feature embedding vectors, and using SiLU activation function to enhance the nonlinear representation of feature embedding vectors.

[0023] Preferably, before extracting the characteristic map of sulcus curvature, cortical gray matter thickness, sulcus depth and myelin sheath in the brain image, the brain image is also preprocessed, specifically including: head motion correction, registration, non-brain tissue removal, gray-white matter boundary reconstruction and deviation correction.

[0024] The present invention also provides a premature infant language development assessment system, comprising:

[0025] A data acquisition module is used to acquire a brain image of a newborn, extract feature maps of sulcus curvature, cortical gray matter thickness, sulcus depth, and myelin sheath in the brain image, divide the brain image into multiple triangular image blocks, and use the feature maps of sulcus curvature, cortical gray matter thickness, sulcus depth, and myelin sheath as feature channels of each vertex of each triangular image block to obtain feature embedding vectors of the left and right hemispheres;

[0026] The language development assessment module is used to input the feature embedding vectors of the left and right brain into a bidirectional scanning network, perform forward and reverse processing on the input data, obtain the nonlinear features between the feature embedding vectors of the left and right brain, use a state space model to compress the nonlinear features into a state space of a fixed size and output them, and evaluate the language development of premature infants based on the output data. The present invention also provides a computer device, including a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the premature infant language development assessment method.

[0027] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the method for assessing language development of premature infants.

[0028] The method for assessing language development of premature infants provided by the present invention has the following beneficial effects:

[0029] The present invention divides the left and right hemispheres of a brain image into multiple triangular image blocks, and uses the feature maps of sulcus curvature, cortical gray matter thickness, sulcus depth and myelin sheath as feature channels for each vertex of each triangular image block to obtain feature embedding vectors for the left and right hemispheres. This process can better capture the geometric features and connection patterns of the brain; by inputting the obtained feature embedding vectors into a bidirectional scanning network, the parameters of the bidirectional scanning network can be made dependent on the input feature embedding vectors; by using a state space model, the input information can be compressed into a state space of a fixed size, irrelevant information can be filtered out, and the focus can be selectively placed on what must be focused on, rather than storing its key-value pairs, thereby effectively reducing existing demands, reducing computational complexity, and improving the efficiency of language development assessment for premature infants. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiment of the present invention and its design scheme, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0031] Figure 1 is a flow chart of a method for assessing language development of premature infants according to an embodiment of the present invention;

[0032] Figure 2 is a detailed flow chart of a method for assessing language development of premature infants according to an embodiment of the present invention;

[0033] Figure 3 is the efficiency comparison between SiT and SiM models, where Figure 3 (a) is the speed comparison result, Figure 3 (b) is the GPU memory comparison result;

[0034] Figure 4 For the comparison between predicted language scores and actual language scores. DETAILED DESCRIPTION

[0035] In order to enable those skilled in the art to better understand the technical solution of the present invention and implement it, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the scope of protection of the present invention.

[0036] In the description of the present invention, it is to be understood that the terms “center”, “longitudinal”, “lateral”, “length”, “width”, “thickness”, “up”, “down”, “front”, “back”, “left”, “right”, “vertical”, “horizontal”, “top”, “bottom”, “inside”, “outside”, “axial”, “radial”, “circumferential”, etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the technical solutions of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0037] In addition, the terms "first", "second", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance. In the description of the present invention, it should be noted that, unless otherwise clearly specified or limited, the terms "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. In the description of the present invention, unless otherwise specified, "plurality" means two or more, which will not be described in detail here.

[0038] Example

[0039] The present invention provides a method for evaluating language development of premature infants, specifically Figure 1 and Figure 2 As shown, the following steps are included:

[0040] Step 1: Obtain a newborn's brain image, extract the feature maps of sulcus curvature, cortical gray matter thickness, sulcus depth, and myelin sheath in the brain image, divide the brain image into multiple triangular image blocks, and use the feature maps of sulcus curvature, cortical gray matter thickness, sulcus depth, and myelin sheath as the feature channels of each vertex of each triangular image block to obtain the feature embedding vectors of the left and right hemispheres.

[0041] The brain image data used in the present invention is the magnetic resonance imaging data (MRI images) of the brain structure of newborns. It is required that all acquisition parameters remain consistent when acquiring the image data, and the brain image analysis software FreeSurfer is used to preprocess the original image data. The preprocessing steps included head motion correction, registration, non-brain tissue removal, reconstruction of gray-white matter boundaries, and bias correction. The Draw-EM algorithm (Developing brain Region Annotation With Expectation-Maximization), a neonatal brain tissue segmentation algorithm, was then used to segment the brain images into tissue types such as cerebrospinal fluid, white matter, cortical gray matter, and subcortical gray matter. Next, the pial surface was generated by extending the white matter grid toward the gray matter and cerebrospinal fluid boundaries, and the sulcal curvature was estimated from the white matter surface. The gray matter cortical thickness was estimated based on the Euclidean distance between the white matter surface and the pial surface. An expansion surface was generated by performing expansion-based smoothing on the white matter surface, and the sulcal depth was estimated during expansion. The myelin map was generated using the ratio of weighted images T1w (T1-weighted imaging) and weighted images T2w (T2-weighted imaging) using FreeSurfer software. Among them, T1w and T2w are two commonly used types in MRI images. T1w is mainly used to study the anatomical structure of the brain, while T2w helps to observe areas with higher water content and is usually used to detect brain lesions. FreeSurfer usually processes T1w to reconstruct the cortex and subcortical areas. Using the ratio of T1w and T2w can help distinguish between myelinated areas (myelin generally shows higher signals in T1-weighted images) and other brain tissues, thereby generating a myelin feature map.

[0042] After extracting the characteristic maps of sulcus curvature, cortical thickness, depth and myelin sheath, the spherical surface of the left and right hemispheres is resampled to a sixth-order regular icosahedron and divided into 1280 triangular meshes, each of which contains 45 vertices. The present invention represents the left and right brain features of each sample as follows:

[0043] Wherein, N=1280, V=45, C=4;

[0044] in, and Represent the original input data of the left and right hemispheres, N represents the number of triangular image blocks in each hemisphere, V is the number of vertices in a triangular image block, and C represents the number of channels of the feature, including sulcus depth, mean curvature, cortical thickness, and myelin sheath (each vertex contains these 4 channels). Represents a set of real numbers, here a three-dimensional tensor or matrix with dimensions N×V×C.

[0045] Afterwards, Flatten to get And sent to the multi-layer perceptron MLP (Multilayer Perceptron) to map into a feature embedding vector of dimension D. A learnable class token is spliced ​​between the left and right hemisphere feature embedding vectors and a one-dimensional position encoding is added to retain the position information, that is:

[0046]

[0047] in, As input to the model of the present invention, and Represent the first triangular image blocks of the left and right hemispheres respectively, W is a linear layer, and are the original input data, with dimensions of N×V×C, L and R are and The flattened data has a dimension of N×(VC), E pos represents the position code, X cls is a learnable embedding vector.

[0048] Step 2: Input the feature embedding vectors of the left and right hemispheres into the bidirectional scanning network, perform forward and reverse processing on the input data, obtain the nonlinear features between the feature embedding vectors of the left and right hemispheres, use the state space model to compress the nonlinear features into a state space of a fixed size and output them, and evaluate the language development of premature infants based on the output data. The specific process is as follows:

[0049] 1) MLP: Input S0 is mapped to dimension E through two different MLPs, and the results are and

[0050] 2) Then, x is processed from the forward and reverse directions respectively. For each direction, a one-dimensional convolution and SiLU activation function are first applied to x. The one-dimensional convolution is used to establish the local contextual relationship between feature embeddings, while the SiLU activation function helps to enhance the nonlinear representation and make the model converge faster. It can be expressed as x' o =SiLU(Conv1d o (x)), where o is forward or backward. Next, x' o Linear mapping to B o , C o and Δ o :

[0051]

[0052] Discretize the parameters A and B using Δ:

[0053]

[0054] Among them, A is the evolution parameter, B is the input parameter, and C is the projection parameter.

[0055] 3) State space model: In order to model the internal connections between different brain regions, the present invention introduces a state space model, which is used to process the feature embedding of these different regions and capture the implicit long-distance dependencies therein, complementing the previous one-dimensional convolution. Specifically:

[0056]

[0057] Among them, h o and o Initialize with zero matrix, i loops from 0 to 2N, and gets y forward and backward ,y forward and backward are the outputs of the forward state space model and the backward state space model respectively, and then use z to y forward and backward To perform gating:

[0058] y' forward =y forward ⊙SiLU(z);

[0059] y' backward =y backward ⊙SiLU(z);

[0060] The two are added together and fed into an MLP, and then added to the initial input S0 to get the output S1:

[0061] S1=Lienear S (y' forward +y' backward )+s0;

[0062] Since the model contains 24 layers (i.e. Figure 2 The output of the first layer of the bidirectional scanning network is also the input of the first layer of the network. The category symbol in the middle of the last layer of the output sequence is input into the MLP to predict the language development score of the newborn. The calculation process is:

[0063]

[0064] in, The vector of the intermediate position of the last layer output.

[0065] The parameters in the above model are all obtained through training. The trained model is used to evaluate the language development of newborns. The difference between the predicted language score and the actual language score is used to accurately evaluate the current language development level and damage of the newborn, which facilitates early intervention.

[0066] In order to verify the performance of the method of the present invention, the method of the present invention (SiM) is compared with the performance of other models. Table 1 lists the comparison results of the SiM model of the present invention with the benchmark geometric deep learning GDL model (Geometric Deep Learning) and the model based on the attention mechanism, wherein Supervised means training from scratch, Fine-tuning means using ImageNet pre-trained weights as initialization parameters and fine-tuning, SiM's Tiny, Small and Base correspond to the three model variants in VisionMamba (Vim), "OOM" means video memory overflow, the GDL models here include MoNet, S2CNN, ChebNet, GConvNet, PointNet++, Spherical UNet, and the training strategy includes training from scratch and fine-tuning using ImageNet pre-trained weights. It is obvious that the results of all model variants of the present invention are significantly higher than those of other models. It is worth noting that when the model parameters and the number of multiplication and accumulation operations MACs (Multiply-Accumulate Operations) are similar, SiM-Small has the best performance, with a MAE of 2.98±2.12.

[0067] Table 1 Performance comparison results of SiM and other models

[0068]

[0069] Figure 3 The efficiency of SiT and SiM is compared in Figure 3 As can be seen from (a), the batch reasoning speed of SiM-Small is 4.7 times faster than that of SiT-Small, and the batch reasoning speed of SiM-Base is 4.6 times faster than that of SiT-Base. Figure 3 As can be seen from (b), SiM-Base saves 50.5% of GPU memory compared to SiT-Base. In practical applications, SiM can significantly improve diagnostic efficiency, helping doctors obtain accurate results in a short time, while reducing dependence on high-end hardware and reducing equipment costs. Higher computing efficiency allows hospitals to process more patient data at the same time, shortening waiting time.

[0070] The best performing model (SiM-Small trained from scratch) was applied to a separate dataset of premature newborns, and the predicted language scores were compared with the actual language scores. Figure 4 As shown, it can be seen that the predicted language score is highly correlated with the actual language score (Pearson correlation coefficient r = 0.93, significance level of Pearson correlation test p < 0.0001), indicating that the model of the present invention can accurately predict the language score of premature newborns, achieve early diagnosis, and help take targeted intervention measures. Improve the quality of medical services. In addition, such models are easy to integrate into portable devices, promoting the development of telemedicine and bedside diagnosis.

[0071] Figure 4 In order to use the present invention to evaluate the language development of premature infants at 18 months, the graph can be used to accurately evaluate the brain magnetic resonance imaging of newborns, understand the language development status of premature infants, and help clinically carry out targeted early intervention.

[0072] The present invention also provides a premature infant language development assessment system, comprising:

[0073] A data acquisition module is used to acquire a brain image of a newborn, extract feature maps of sulcus curvature, cortical gray matter thickness, sulcus depth, and myelin sheath in the brain image, divide the brain image into multiple triangular image blocks, and use the feature maps of sulcus curvature, cortical gray matter thickness, sulcus depth, and myelin sheath as feature channels of each vertex of each triangular image block to obtain feature embedding vectors of the left and right hemispheres;

[0074] The language development assessment module is used to input the feature embedding vectors of the left and right brain into a bidirectional scanning network, perform forward and reverse processing on the input data, obtain the nonlinear features between the feature embedding vectors of the left and right brain, use the state space model to compress the nonlinear features into a state space of a fixed size and output them, and assess the language development of premature infants based on the output data.

[0075] The present invention also provides a computer device, including a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute a method for evaluating the language development of premature infants.

[0076] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program is suitable for being loaded by a processor to execute a method for evaluating language development of premature infants.

[0077] The embodiments described above are only preferred specific implementation modes of the present invention, and the protection scope of the present invention is not limited thereto. Any simple changes or equivalent replacements of the technical solutions that can be obviously obtained by any technician familiar with the field within the technical scope disclosed in the present invention belong to the protection scope of the present invention.

Claims

1. A method for assessing language development of premature infants, characterized in that: The steps include: Obtain a newborn's brain image, extract the feature map of sulcus curvature, cortical gray matter thickness, sulcus depth and myelin sheath in the brain image, divide the brain image into multiple triangular image blocks, and use the feature map of sulcus curvature, cortical gray matter thickness, sulcus depth and myelin sheath as the feature channel of each vertex of each triangular image block to obtain the feature embedding vectors of the left and right hemispheres; The feature embedding vectors of the left and right brain are input into a bidirectional scanning network, and the input data is forward processed and reverse processed to obtain the nonlinear features between the feature embedding vectors of the left and right brain. The state space model is used to compress the nonlinear features into a state space of a fixed size and output them. The language development of premature infants is evaluated based on the output data.

2. The method for assessing language development of premature infants according to claim 1, characterized in that: The method of extracting the characteristic map of sulcus curvature, cortical gray matter thickness, sulcus depth and myelin sheath in the brain image specifically comprises the following steps: The Draw-EM algorithm was used to segment brain images into cerebrospinal fluid, white matter, cortical gray matter, and subcortical gray matter; The pial surface was generated by extending the white matter mesh toward the cortical gray matter and cerebrospinal fluid boundaries, and the sulcal and gyral curvature was estimated from the white matter surface; Cortical gray matter thickness was estimated based on the Euclidean distance between the white matter surface and the pial surface; Dilated surfaces were generated by performing extension-based smoothing of the white matter surface, and sulcus-gyrus depths were estimated during dilation; The ratio of weighted T1w and weighted T2w images was used to generate a characteristic map of myelin sheaths using FreeSurfer software.

3. The method for assessing language development of premature infants according to claim 1, characterized in that: The step of obtaining the feature embedding vectors of the left and right hemispheres of the brain comprises the following steps: The left and right brain surfaces of the brain image were resampled onto a sixth-order regular icosahedron, and the left and right brain surfaces were divided into 1280 triangular image blocks, each containing 45 vertices. The feature maps of sulcus curvature, cortical gray matter thickness, sulcus depth and myelin sheath are used as feature channels of each vertex of each triangular image block to obtain the feature representation of the left and right hemispheres. The feature representations of the left and right hemispheres are flattened and input into the MLP to map them into feature embedding vectors of dimension D.

4. The method for assessing language development of premature infants according to claim 1, characterized in that: Before inputting the feature embedding vectors of the left brain and the right brain into the bidirectional scanning network, the method also includes splicing a learnable category symbol between the feature embedding vector of the left brain and the feature embedding vector of the right brain and adding a one-dimensional position encoding.

5. The method for assessing language development of premature infants according to claim 1, characterized in that: The bidirectional scanning network includes 24 layers, each of which includes two MLPs, two one-dimensional convolutions after one of the MLPs, two SiLU activation functions, two state-space models, and a SiLU activation function after another MLP.

6. The method for assessing language development of premature infants according to claim 5, characterized in that: The input data is forward processed and reverse processed, specifically: using one-dimensional convolution in each direction to establish a local contextual relationship between feature embedding vectors, and using the SiLU activation function to enhance the nonlinear representation of the feature embedding vectors.

7. The method for assessing language development of premature infants according to claim 1, characterized in that: Before extracting the characteristic map of sulcus curvature, cortical gray matter thickness, sulcus depth and myelin sheath in the brain image, the brain image is also preprocessed, specifically including: head motion correction, registration, non-brain tissue removal, gray-white matter boundary reconstruction and deviation correction.

8. A system for assessing language development of premature infants, characterized in that: include: A data acquisition module is used to acquire a brain image of a newborn, extract feature maps of sulcus curvature, cortical gray matter thickness, sulcus depth, and myelin sheath in the brain image, divide the brain image into multiple triangular image blocks, and use the feature maps of sulcus curvature, cortical gray matter thickness, sulcus depth, and myelin sheath as feature channels of each vertex of each triangular image block to obtain feature embedding vectors of the left and right hemispheres; The language development assessment module is used to input the feature embedding vectors of the left and right brain into a bidirectional scanning network, perform forward and reverse processing on the input data, obtain the nonlinear features between the feature embedding vectors of the left and right brain, use the state space model to compress the nonlinear features into a state space of a fixed size and output them, and assess the language development of premature infants based on the output data.

9. A computer device, characterized in that: It comprises a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to execute the method for assessing language development of premature infants as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the method for assessing language development of premature infants as described in any one of claims 1-7.

Images