ES2635285B2 - Method and system for non-invasive characterization of human and animal tissues in vivo - Google Patents

Description

DESCRIPTION

Method and system for non-invasive characterization of human and animal tissues in vivo.

Sector of the technique

The present invention fits into the field of methods and apparatus for determining the composition of animal or human tissues in vivo non-invasively. More specifically, the invention relates to the intelligent determination of the breast tissue composition by combining spectroscopic techniques and neural networks.

State of the art

Raman spectroscopy is a photonic technique capable of providing molecular information of materials, whether organic or inorganic, allowing its identification.

The analysis by means of Raman spectroscopy is based on the treatment of light scattered by a material when striking a monochromatic beam of light. A small portion of the incident light is dispersed inelastically, experiencing slight changes in frequency that are characteristic of the chemical structure of the material analyzed, and independent of the frequency of the incident light. It is an analysis technique that is performed directly on the material to be analyzed without needing any special type of preparation and that does not involve any alteration of the surface on which the analysis is performed; that is, it is a non-destructive technique. For this reason it has been applied in the analysis of a great variety of biological tissues such as fluids, cells and tissues.

The biochemical changes in cells and tissues, which can be caused or be the cause of diseases, can lead to significant changes in the Raman spectrum, so this technique presents a high potential for use in the diagnosis and prognosis of different diseases. as a tool to evaluate new therapies (K. Kong et al, Advanced Drug Delivery Reviews 89 (2015) 121-134).

Raman spectroscopy has been applied in breast cancer detection identifying microcalcifications in malignant and benign lesions. Microcalcifications are markers of cellular degradation and necrosis and are chemically composed of calcium oxalate and hydroxyapatite mineral deposits (Radi MJ., Axch Pathol Lab Med 113 (1989) 1367-9). In general, calcium oxalate is associated with benign lesions while hydroxyapatite deposits are associated with proliferative lesions such as carcinoma. Based on the analysis of these calcium compounds, Haka et al. (Cancer Research 62 (2002) 5375-5380) concluded that the Raman technique can be applied, firstly, in-vitro on small samples of biopsied tissue being desirable its in-vivo application combined with the mammography to select patients with microcalcifications that They need to have a biopsy.

On the other hand, changes in tissue composition associated with malignancy are reflected not only in the composition of microcalcifications but also in an increase in nucleoproteins and nucleic acids, so a complete biochemical analysis could lead to more reliable results.

The in-vitro application on biopsied tissue, in addition to being an invasive technique, has the disadvantage that a small sample of tissue is analyzed, so it can happen that the sample is not representative of the entire breast volume. The in-vivo application has the disadvantage that it is only able to provide very close tissue information to the surface (few tens of micrometers deep). To increase the depth of the analyzed tissue that allows a biochemical exploration of a large area (as in the case of the breast), it is necessary a time of analysis so high that the technique is not viable in practice.

To overcome this drawback in US20090238333 a method is described where tissue compositions are analyzed in different areas of the outer surface thereof and, by means of a numerical model, the composition of the rest of tissue is predicted. This method allows to know the real composition of the tissue in superficial layers but the composition in deeper layers is only an estimate that results from the application of a mathematical method so it does not allow, therefore, to know with certainty the real composition of the tissue in deep layers.

Therefore, it would be desirable to develop an in-vivo, non-invasive study method capable of reliably analyzing the chemical composition of breast tissue in both superficial and deep layers and at a time that makes it possible to use it in the diagnosis and monitoring of the breast cancer Likewise, it would be necessary to include intelligent algorithms to be able to consider the changes in the analytes of the chemical composition and be applied to more efficient diagnosis and monitoring of breast cancer.

Description of the invention

In the present invention, a method of in vivo analysis of animal or human tissue is described to determine the chemical composition and characteristics thereof that combines a fast scanning of diffuse optics and a precise Raman scan allowing the biochemical exploration of a large area in short. space of time.

The identification of the type of tissue according to its instantaneous chemical composition is possible thanks to a spectral library obtained by a laser pulse in only one of the representative samples and its subsequent treatment using neural networks. The results obtained show an efficacy of more than 99.99% in the tissues analyzed.

The method combines:

- Infrared low-precision diffuse optical scanning (DOT), whose error is tilted to the excess of false positives, in order to determine areas suspected of containing certain chemical compounds. Storage of scans of suspicious areas of information in neural networks.

- Analysis of the data obtained in the DOT infrared scan through the use of designed neural networks and the previously created neuronal database for the classification of the suspect tissue.

- Accurate scanning Raman / SORS (Spatially Offset Raman Spectroscopy) bounded to areas of suspicious tissue. Storage of accurate and bounded scans of suspicious areas in neural networks.

- Analysis of the data obtained in precise Raman / SORS scanning through the use of trained neural networks with known Raman signatures database of previous Raman scans and stored information.

- Determination of the chemical composition and / or characteristics of the tissue by means of software that combines the results of the neural networks of diffuse optics and Raman analysis with specific criteria such as, for example, the morphology of the calcifications present in the tissue.

The first scan is performed to find areas where, subsequently, apply Raman spectroscopy and consists of a diffuse optical tomography (DOT) with near infrared beam (NIR, Near InfraRed). The information obtained with a NIR beam of light is treated raw to feed a real-time neural network that determines if a Raman exploration is necessary in a specific area. Likewise, this information of conflicting zones is stored with its chemical composition collected in the neural networks to improve the classifications over time. The transport of light in tissue at these wavelengths in a certain direction of dispersion becomes almost isotropic, so it can be predicted correctly with a diffuse optics model. For the analysis of the data obtained by fuzzy optics, a finite element model and image reconstruction algorithms are used.

To the information obtained in the first scan by DOT, the analysis of the Raman bands of a second scan by Raman spectrometry is added.

Raman scanning is performed using the spatial compensation technique Raman SORS that is effective at depths of 500 mm in biological tissue. The Raman signal is coupled through an optical image projection to an optical fiber beam and delivered to a high performance infrared (Charge Coupled Device) CCD sensor. In this way, detailed information on the chemical composition of the tissues is obtained. Given the great variability that appears in cancers and chemical compositions, the creation of databases in neural networks give the possibility of resolving spectra with much more reliability. Asl, for example, healthy mammary tissue in humans shows Raman bands at 1078, 1300, 1445 and 1651 cm-1 and malignant tissue only shows bands at 1445 and 1651 cm-1. Benign tumors have characteristic bands at 1240, 1445 and 1659 cm-1. In addition, the intensities of bands 1445 and 1651 cm-1 are correlated with the classification of the disease.

The obtained Raman data is analyzed using a neural network (NN) with forward connections, specifically following a propagation perceptron model (retropropagation, perceptron model). The NN consists of three layers called input, output and hidden, formed by neurons (only operative element). The input layer is used only for the entry of the data matrix in the NN. In the other layers, non-linear calculations are made. In the hidden layer each neuron receives signals from other input nodes, adding these through the activation function. Then the result is transformed by the transfer function to subsequently be sent to the output neurons to determine if the recognition is positive or negative. All these neurons and nodes are linked forward and each union is weighted by means of variables called weights. These weights are responsible for adapting to the system to model and "remember" the information presented to the model. The basic backpropagation algorithms adjust the weights in a steep slope of descending direction (negative values of the gradient); that is, the direction in which the performance function decreases more rapidly. The basic measurements are the number of positives and negatives (true and false, PV, NV, FP, FN) from which the sensitivity and specificity of the detection processes is determined. A true positive (PV) corresponds to the correct detection of a substance, compound or characteristic in a sample, when it actually exists. A true negative (NV) corresponds to the negative detection of a substance, compound or characteristic of a sample when it does not exist. The detections are considered false (PF, NF) when the detection does not correspond to the reality of the sample.

The obtained Raman data is stored in a neural network (NN) for later analysis and to improve the resolutions of the spectra. This storage is carried out by medium of a neural network and its weights. These weights are adapted to each of the data of conflicting spectra and their chemical composition presented to the network. In this way, the system under study and its composition are recorded in the form of a numerical matrix that adapts to the new spectra / composition that are presented over time.

The interaction of light with matter in a linear regime allows the absorption and emission of light that adjusts to the energy levels already defined by the electrons. The Raman effect corresponds, in the theory of perturbations of quantum mechanics, to the absorption and consequent emission of a photon by changing the intermediate state of an electron, passing through a virtual state. There are the following possibilities:

1. There is no energy exchange between the incident photons and the molecules (there is, therefore, no Raman effect)

2. There is energy exchange between the incident photons and the molecules. The energy differences are equal to the differences of the vibrational or rotational states of the molecule. These are differences in energy measured by subtracting the energy of a mono-energetic laser from light scattered photons. In crystals, only certain photons are admitted by the crystal structure so that the Raman scattering effect can appear only at certain frequencies.

• Molecules absorb energy (Stokes scattering). The resulting photon is of lower frequency and generates a Stokes line on the red side of the incident spectrum.

• The molecule loses energy (anti-Stokes scattering). The incident photons are displaced at higher frequencies (blue) of the spectrum and, therefore, generate a line called anti-Stokes.

The intensities of the Raman bands only depend on the number of molecules that occupy the different vibrational states; The variation of the frequencies determines the chemical structure of the material analyzed.

Mode of realization of the invention

The present invention is illustrated by the following examples, which are not intended to be limiting of the scope thereof.

Example 1

This example refers to the infrared preliminary scan and creation of the intelligent database of conflicting data.

In the first scan to find areas to apply Raman spectroscopy, a diffuse optics tomography with infrared beam is performed. The information derived from the exploration with a NIR light beam of wavelengths between 650 and 900 nm is treated raw to feed the real-time neural network so that the intrinsic optical properties of the tissue can be extracted in each of the different wavelengths applied, thus obtaining! the amount of hemoglobin or the proportion of water and the saturation of oxygen in each tissue area. The results obtained can also be interpreted graphically to distinguish between tissue areas that are suspicious and not suspicious of presenting a lesion and on which to perform a subsequent analysis using Raman SORS. This information of optical properties and concentration of hemoglobin or water from the conflictive zones are included in a neural network to improve the subsequent analyzes.

Example 2

This example refers to the analysis of breast tissue using the Raman SORS technique. Creation of the database in the form of a neural network.

A Raman spectrum is obtained from a sample of breast tissue following the SORS method described in US7652763 and using a laser diode for Raman spectroscopy with temperature stabilization, attenuating at 115 mW and operating at 827 nm (Micro Laser Systems, Lepton IV Series Diffraction Limited Diode Lasers L4-830S). The power incident on the sample is 50 mW with a beam diameter of 0.5 -1 mm. The beam is spectrally filtered to remove spontaneous emission components using two 830 nm band pass filters from Semrock. The beam impinges on the sample at an angle of 45 °. The spatial displacement is 3 mm.

For the acquisition of the Raman light, a lens with a diameter of 50 mm and a focal length of 60 mm is used. The scattered light is collimated and passed through a 50 mm diameter holographic filter ( Kaiser Optical Systems, Inc. Holographic Notch Filter 830 nm ) to remove elastic dispersion components. The Raman light is propagated through SORS annular fiber systems of ~ 2 m in length to the input of the HoloSpec ™ f / 1.8 (.Kaiser Optical Systems, Inc.) spectrograph. The Raman spectrum is obtained through a CCD camera Andor Technology DU420A-BR-DD, 1024 x 256 pixels. The ring fiber system consists of a beam of 50 active fibers responsible for collecting light (scattered by the sample) placed in a circle at a distance of 3 mm around the emission point that directs the laser light towards the sample. By using this configuration the intensity of the signal can be greatly increased.

With exposures of 0.2 to 10 seconds, the spectrum of a volume of 50x50x14 mm is acquired, resulting in the generation of 20 GB of data with which the Raman analysis stage of the neural network is fed. These generated data were transformed into a numeric matrix of weights that store the information gathered in the experiment, reducing its size to 500MB.

Example 3

In this example, the sorting mode of the analyzed tissue is shown.

Using Raman spectrometry, detailed information on the chemical composition of the tissue is obtained. Looking for differences in the Raman readings in mammary tissues it is observed that the healthy tissue shows Raman bands in 1078, 1300, 1445 and 1651 cm-1. Benign tumors have characteristic bands at 1240.1445 and 1659 cm-1. In addition, the intensities relative to the bands 1445 and 1651 cm-1 are correlated with the tissue classification in malignant, benign or normal (Manoharan, R. et al .. Photochemistry and Photobiology, 67 (1), (1998), 15 22 ).

Normal breast tissue is easy to distinguish because it has strong Raman signals from abundance of lipids. However, the differentiation between tissue with malignant or benign lesions is difficult because in both cases there is a high protein content due to stromal changes (Mahoran, R. et al., Photochemistry and Photobiology, 67 (1), (1998)). , 15-22).

To obtain the differentiation, the obtained Raman data are analyzed using a feed-forward neural network, specifically with a multilayer perceptron propagation model. Using retropropagation algorithms for the optimization of weights, as mentioned above.

The main components are analyzed to examine the spectral differences between malignant and benign tumors. Instead of constructing an algorithm based on a series of spectral characteristics, the principal component analysis allows the full range of data to be used to develop a decision algorithm by logistic regression. Using the adjustment coefficients of these main components, we estimate the probability that each sample belongs to a category (normal, benign or malignant).

These results show that the Raman spectra contain the necessary information to differentiate normal, benign and malignant mammary tissue.

Claims (5)

1. Method to characterize breast tissue comprising the following stages: - Infrared scanning of low-precision diffuse optics (DOT) in order to determine areas suspected of containing certain chemical compounds. Storage of scans of suspicious areas of information in neural networks. - Analysis of the data obtained in DOT infrared scanning through the use of designed neural networks and the previously created neuronal database for tissue classification in suspect and non-suspect. - Raman / SORS precise scanning bounded to areas of suspicious tissue. Storage of accurate and bounded scans of suspicious areas in neural networks. - Analysis of the data obtained in precise Raman / SORS scanning through the use of trained neural networks with known Raman signatures database of previous Raman scans and stored information. - Determination of the presence of chemical composition and / or tissue characteristics by software that combines the results of the neural networks of diffuse optics and Raman analysis with specific criteria such as, for example, the morphology of the calcifications present in the tissue. Characterized because the neural network that analyzes the Raman data obtained is a neural network (NN) with forward connections, specifically following a propagation perceptron model (retropropagation, perception model). The NN consists of three layers called input, output and hidden, formed by neurons (only operative element). The input layer is used only for the entry of the data matrix in the NN. In the other layers, non-linear calculations are made. In the hidden layer each neuron receives signals from other input nodes, adding these through the activation function. Then the result is transformed by the transfer function to subsequently be sent to the output neurons to determine if the recognition is positive or negative. All these neurons and nodes are linked forward and each union is weighted by means of variables called weights. These weights are responsible for adapting to the system to model and "remember" the information presented to the model. The basic backpropagation algorithms adjust the weights in a steep slope of descending direction (negative values of the gradient); that is, the direction in which the performance function decreases more rapidly. The basic measurements are the number of positives and negatives (true and false, PV, NV, FP, FN) from which the sensitivity and specificity of the detection processes is determined. A true positive (PV) corresponds to the correct detection of a substance, compound or characteristic in a sample, when it actually exists. A true negative (NV) corresponds to the negative detection of a substance, compound or characteristic of a sample when it does not exist. The detections are considered false (PF, NF) when the detection does not correspond to the reality of the sample. The obtained Raman data is stored in a neural network (NN) for later analysis and to improve the resolutions of the spectra. This storage is carried out by means of a neural network and its weights. These weights are adapted to each of the data of conflicting spectra and their chemical composition presented to the network. In this way, the system object of study and its composition in the form of a numerical matrix that adapts to the new spectra / composition that are presented over time and results are obtained with an efficiency of more than 99.99%. in short space of time and results are obtained with an efficiency of more than 99.99% in a short space of time. 2. Method according to claim 1, wherein the neural network for analyzing data obtained in infrared scanning DOT is previously trained with known DOT data to classify zones in the tissue. 3. Method according to claims 1 and 2, wherein the neural network for analysis of data obtained in the Raman scan is previously trained with known Raman scan data to determine the chemical composition and the characteristics of the tissue. Method, according to previous claims, where the neural network of diffuse optics is fed back with data obtained by diffuse optics in order to carry out a continuous training. 5. Method, according to previous claims, where the neural network of the Raman scan is fed back with data obtained by Raman scanning in order to perform a continuous training.