Spectroscopic detection of cervical pre-cancer using radial basis function networks

• US 6,135,965 A
• Filed: 12/02/1996
• Issued: 10/24/2000
• Est. Priority Date: 12/02/1996
• Status: Expired due to Fees

• Current Assignee: Board of Regents of the University of Texas System (University of Texas System)
• Original Assignee: Board of Regents of the University of Texas System (University of Texas System)
• Inventors: Tumer, Kagan, Ghosh, Joydeep, Ramanujam, Nirmala, Richards-Kortum, Rebecca
• Primary Examiner(s): Lateef, Marvin M.
• Assistant Examiner(s): Shaw, Shawna J.
• Application Number:
  US08/757,116
• Time in Patent Office: 1,422 Days
• Field of Search: 600/407, 600/408, 600/473, 600/475, 600/476, 600/477, 600/478, 600/310, 600/342, 395/21, 395/924, 436/63, 436/64, 436/813, 250/461.2, 128/925
• US Class Current: 600/476
• CPC Class Codes: A61B 5/0071 by measuring fluorescence emission
• CPC Class Codes: A61B 5/0075 by spectroscopy, i.e. measuring spectra, e.g. Raman spectroscopy, infrared absorption spectroscopy A61B5/0071 takes precedence
• CPC Class Codes: A61B 5/0084 for introduction into the body, e.g. by catheters A61B5/1459 takes precedence
• CPC Class Codes: A61B 5/7267 involving training the classification device
• CPC Class Codes: G01J 2003/2866 Markers; Calibrating of scan
• CPC Class Codes: G16H 50/70 for mining of medical data, e.g. analysing previous cases of other patients
• CPC Class Codes: Y10S 128/925 Neural network

Abstract

An apparatus and methods for spectroscopic detection of tissue abnormality, particularly precancerous cervical tissue, using neural networks to analyze in vivo measurements of fluorescence spectra. The invention excites fluorescence intensity spectra in both normal and abnormal tissue. This fluorescence spectroscopy data is used to train a group (ensemble) of neural networks, preferably radial basis function (RBF) neural networks. Once trained, fluorescence spectroscopy data from unknown tissue samples is classified by the trained neural networks. This process is used to differentiate pre-cancers from normal tissues, and can also be used to differentiate high grade pre-cancers from low grade pre-cancers. One embodiment of the invention is able to distinguish pre-cancerous tissue from both normal squamous tissue (NS) and normal columnar (NC) tissue in a single-stage of analysis. The invention demonstrates significantly smaller variability in classification accuracy, resulting in more reliable classification, with superior sensitivity. Moreover, the single-stage embodiment of the invention simplifies the decision-making process as compared to a two-stage embodiment.

Claims

1. An apparatus for detecting and classifying tissue abnormality at a tissue site, comprising:

(a) at least one source of electromagnetic radiation of selected wavelengths that excite different fluorescence intensity spectra in normal and abnormal tissue;

(b) a receiver sensitive to the fluorescence intensity spectra;

(c) a tissue site probe coupled to each source and to the receiver; and

(d) an ensemble of neural networks, coupled to the receiver, for calculating from the fluorescence intensity spectra a probability that the tissue site is normal or abnormal,

wherein each neural network in such ensemble comprises a radial basis function (RBF) network for generating an associated probability estimate and further including a means for combining the probability estimates into a single probability.

2. An apparatus as in claim 1, wherein the means for combining utilizes a median class order statistical combiner.

3. An apparatus as in claim 1, wherein each neural network in such ensemble comprises:

(a) a layer of input processing units receiving an input vector and producing an output;

(b) a layer of hidden processing units each receiving one of the outputs from each of the input processing units and producing an output; and

(c) an output unit receiving each hidden unit output multiplied by a weight, the output unit generating an output that is a function of its inputs.

4. An apparatus as in claim 1, wherein each neural network in such ensemble comprises a multilayer perceptron network for generating a probability estimate, and further including a means for combining the probability estimates into a single probability.

5. The apparatus as in claim 1, further including means for training the ensemble of neural networks using fluorescence intensity spectra from known normal and abnormal tissue.

6. An apparatus as in claim 5, wherein the training means adjusts the weight in an iterative process to produce a desired output in response to a given input, wherein the desired output comprises the probability.

7. An apparatus as in claim 1, wherein the fluorescence intensity spectra derives from abnormal cervical tissue, normal squamous cervical tissue, and normal columnar cervical tissue, wherein the probability is a single probability distinguishing abnormal tissue from both normal squamous and normal columnar tissue.

8. An apparatus as in claim 1, further including means for conducting a principle component analysis of the fluorescence intensity spectra.

9. An apparatus as in claim 8, further including means for normalizing the first fluorescence intensity spectra relative to respective maximum intensities thereof, prior to conducting the principle component analysis.

10. An apparatus as in claim 9, further including means for mean-scaling the first fluorescence intensity spectra as a function of a mean intensity thereof, prior to conducting the principle component analysis.

11. An apparatus as in claim 8, further including means for selecting component loadings calculated from the principal component analysis to reduce the number of excitation-emission wavelength pairs from the fluorescence intensity spectra required to train the ensemble of neural networks.

12. An apparatus as in claim 1, wherein at least one source of electromagnetic radiation comprises a laser operated to generate pulses at each wavelength having a power level, pulse duration, and repetition rate that excites the fluorescence intensity spectra in normal and abnormal tissue.

13. An apparatus as in claim 1, wherein the tissue is cervical tissue, and a probability of abnormal tissue indicates a cancerous or pre-cancerous condition.

14. A method for detecting and classifying tissue abnormality at a tissue site, comprising the steps of:

(a) exciting different fluorescence intensity spectra in normal and abnormal tissue;

(b) receiving the fluorescence intensity spectra; and

(c) calculating from the fluorescence intensity spectra, using an ensemble of neural networks, a probability that the tissue site is normal or abnormal,

wherein each neural network in such ensemble comprises a radial basis function (RBF) network for generating an associated probability estimate, and further including the step of combining the probability estimates into a single probability.

15. A method as in claim 14, wherein the step of combining utilizes a median class order statistical combiner.

16. A method as in claim 14, wherein each neural network in such ensemble comprises:

(a) a layer of input processing units receiving an input vector and producing an output;

(b) a layer of hidden processing units each receiving one of the outputs from each of the input processing units and producing an output; and

(c) an output unit receiving each hidden unit output multiplied by a weight, the output unit generating an output that is a function of its inputs.

17. A method as in claim 14, each neural network in such ensemble comprises a multilayer perceptron network for generating a probability estimate, and further including the step of combining the probability estimates into a single probability.

18. The method as in claim 14, further including the step of training the ensemble of neural networks using fluorescence intensity spectra from known normal and abnormal tissue.

19. A method as in claim 18, further including the step of adjusting weights in each neural network in an iterative process to produce a desired output in response to a given input, wherein the desired output comprises the probability.

20. A method as in claim 14, wherein the fluorescence intensity spectra derives from abnormal cervical tissue, normal squamous cervical tissue, and normal columnar cervical tissue, wherein the probability is a single probability distinguishing abnormal tissue from both normal squamous and normal columnar tissue.

21. A method as in claim 14, further including the step of conducting a principle component analysis of the fluorescence intensity spectra.

22. A method as in claim 21, further including the step of normalizing the first fluorescence intensity spectra relative to respective maximum intensities thereof, prior to conducting the principle component analysis.

23. A method as in claim 22 further including the step of mean-scaling the first fluorescence intensity spectra as a function of a mean intensity thereof, prior to conducting the principle component analysis.

24. A method as in claim 21, further including the step of selecting component loadings calculated from the principal component analysis to reduce the number of excitation-emission wavelength pairs from the fluorescence intensity spectra required to train the ensemble of neural networks.

25. A method as in claim 14, wherein the different fluorescence intensity spectra are excited by a laser operated to generate electromagnetic radiation at selected wavelengths.

26. A method as in claim 14, wherein the tissue is cervical tissue, and a probability of abnormal tissue indicates a cancerous or pre-cancerous condition.

27. A method for in vivo analysis of cervical tissue, comprising the steps of:

(a) inserting an optical probe within a cervix, the probe having a light source and a light receptor;

(b) illuminating a selected area of the cervix with selected wavelengths of light from the light source;

(c) exciting fluorescence intensity spectra in both normal and abnormal tissue in the cervix with the light;

(d) receiving the fluorescence intensity spectra from the selected area through the light receptor;

(e) analyzing the received fluorescence intensity spectra, using an ensemble of neural networks, to determine a probability that the cervical tissue in the selected area is normal or abnormal

wherein each neural network in such ensemble comprises a radial basis function (RBF) network for generating an associated probability estimate, and further including the step of combining the probability estimates into a single probability.

28. A method as in claim 27, wherein each neural network in such ensemble comprises a multilayer perceptron network for generating a probability estimate, and further including the step of combining the probability estimates into a single probability.

29. A method for analyzing fluorescence intensity spectra from a tissue site in order to detect and classify tissue abnormality at the tissue site, comprising the step of:

(a) calculating from the fluorescence intensity spectra, using an ensemble of neural networks, a probability that the tissue site is normal or abnormal

wherein each neural network in such ensemble comprises a radial basis function (RBF) network for generating an associated probability estimate, and further including the step of combining the probability estimates into a single probability.

30. A method as in claim 29, wherein the step of combining utilizes a median class order statistical combiner.

31. A method as in claim 29, wherein each neural network in such ensemble comprises:

(a) a layer of input processing units receiving an input vector and producing an output;

(b) a layer of hidden processing units each receiving one of the outputs from each of the input processing units and producing an output; and

(c) an output unit receiving each hidden unit output multiplied by a weight, the output unit generating an output that is a function of its inputs.

32. A method as in claim 29, wherein each neural network in such ensemble comprises a multilayer perceptron network for generating a probability estimate, and further including the step of combining the probability estimates into a single probability.

33. The method as in claim 29, further including the step of training the ensemble of neural networks using fluorescence intensity spectra from known normal and abnormal tissue.

34. A method as in claim 33, further including the step of adjusting weights in each neural network in an iterative process to produce a desired output in response to a given input, wherein the desired output comprises the probability.

35. A method as in claim 29, wherein the fluorescence intensity spectra derives from abnormal cervical tissue, normal squamous cervical tissue, and normal columnar cervical tissue, wherein the probability is a single probability distinguishing abnormal tissue from both normal squamous and normal columnar tissue.

36. A method as in claim 29, further including the step of conducting a principle component analysis of the fluorescence intensity spectra.

37. A method as in claim 36, further including the step of normalizing the first fluorescence intensity spectra relative to respective maximum intensities thereof, prior to conducting the principle component analysis.

38. A method as in claim 36, further including the step of mean-scaling the first fluorescence intensity spectra as a function of a mean intensity thereof, prior to conducting the principle component analysis.

39. A method as in claim 36, further including the step of selecting component loadings calculated from the principal component analysis to reduce the number of excitation-emission wavelength pairs from the fluorescence intensity spectra required to train the ensemble of neural networks.

40. A method as in claim 29, wherein the fluorescence intensity spectra are excited by a laser operated to generate electromagnetic radiation at selected wavelengths.

41. A method as in claim 29, wherein the tissue is cervical tissue, and a probability of abnormal tissue indicates a cancerous or pre-cancerous condition.

42. A computer program, residing on a computer-readable medium, for detecting and classifying tissue abnormality at a tissue site using data in a computer derived from fluorescence intensity spectra of normal and abnormal tissue, the computer program comprising instructions for causing a computer to:

(a) pre-process the fluorescence intensity spectra data; and

(b) calculate a probability that the tissue site is normal or abnormal from the fluorescence intensity spectra data using an ensemble of neural networks,

wherein the computer program further comprises instructions for causing the computer to calculate the probability using an ensemble of radial basis function (RBF) networks, each generating an associated probability estimate, and to combine the probability estimates into a single probability.

43. A computer program as in claim 42, wherein the computer program further comprises instructions for causing the computer to train each RBF network using fluorescence intensity spectra from known normal and abnormal tissue.

44. A computer program as in claim 42, wherein the computer program further comprises instructions for causing the computer to conduct a principle component analysis of the fluorescence intensity spectra.

45. A computer program as in claim 44, further including instructions for causing the computer to select component loadings calculated from the principal component analysis to reduce the number of excitation-emission wavelength pairs from the fluorescence intensity spectra required to train the ensemble of neural networks.

46. A computer program as in claim 42, wherein the computer program further comprises instructions for causing the computer to calculate the probability using an ensemble of multilayer perceptron works, each generating an associated probability estimate and to combine the probability estimates into a single probability.