Method and system for differential diagnosis based on clinical and radiological information using artificial neural networks

• US 5,622,171 A
• Filed: 04/14/1995
• Issued: 04/22/1997
• Est. Priority Date: 08/28/1990
• Status: Expired due to Term

• Current Assignee: ARCH Development Corporation (University of Chicago)
• Original Assignee: ARCH Development Corporation (University of Chicago)
• Inventors: Asada, Naoki, Doi, Kunio
• Primary Examiner(s): Smith, Ruth S.
• Application Number:
  US08/422,172
• Time in Patent Office: 739 Days
• Field of Search: 364/413.01, 364/413.02, 364/413.13, 378/901, 128/630, 128/653.1, 128/660.07, 128/915, 395/11, 395/20-23, 395/24
• US Class Current: 600/408
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G16H 30/20 for handling medical images, e.g. DICOM, HL7 or PACS
• CPC Class Codes: G16H 50/20 for computer-aided diagnosis, e.g. based on medical expert systems
• CPC Class Codes: G16H 50/30 for calculating health indices; for individual health risk assessment
• CPC Class Codes: G16H 70/60 relating to pathologies
• CPC Class Codes: Y10S 128/922 including image analysis
• CPC Class Codes: Y10S 128/924 using artificial intelligence
• CPC Class Codes: Y10S 128/925 Neural network

Abstract

A method and system for computer-aided differential diagnosis of diseases, and in particular, computer-aided differential diagnosis using neural networks. A first embodiment of the neural network distinguishes between a plurality of interstitial lung diseases on the basis of inputted clinical parameters and radiographic information. A second embodiment distinguishes between malignant and benign mammographic cases based upon similar inputted clinical and radiographic information. The neural networks were first trained using a hypothetical data base made up of hypothetical cases for each of the interstitial lung diseases and for malignant and benign cases. The performance of the neural network was evaluated using receiver operating characteristics (ROC) analysis. The decision performance of the neural network was compared to experienced radiologists and achieved a high performance comparable to that of the experienced radiologists. The neural network according to the invention can be made up of a single network or a plurality of successive or parallel networks. The neural network according to the invention can also be interfaced to a computer which provides computerized automated lung texture analysis to supply radiographic input data in an objective and automated manner.

Claims

1. A method for differential diagnosis of a plurality of breast diseases, the method comprising the steps of:

selecting a plurality of clinical parameters defining characteristics of a subject;

selecting radiographic descriptors comprising predetermined features obtained from a radiographic breast image defining characteristics of breast diseases, said radiographic descriptors consisting essentially of each of the following twelve radiographic descriptors;

shape of density, size of density, margin of density, margin spiculation and pattern of density, presence of, number of, shape of, uniformity of, and distribution of microcalcifications, parenchymal distortion and skin thickening;

converting said plurality of clinical parameters and said radiographic descriptors into numerical expressions;

transforming each of said numerical expressions into a number in a predetermined range;

inputting said transformed numerical expressions into a neural network; and

diagnosing at least one of said plurality of breast diseases using said neural network in accordance with said input expressions.

2. A method as recited in claim 1, wherein said neural network comprises 3 layers, 45 input units, at least one hidden unit and at least one output unit.

3. A method for differential diagnosis of a plurality of interstitial lung diseases, the method comprising the steps of:

selecting a plurality of clinical parameters defining characteristics of a subject;

selecting radiographic descriptors comprising predetermined features obtained from a radiographic chest image defining characteristics of interstitial lung diseases, said radiographic descriptors consisting essentially of each of the following fourteen radiographic descriptors;

distribution of infiltrates in 6 lung zones, homogeneity, fineness, nodularity, septal lines and honeycombing of said infiltrates, lymphadenopathy, pleural effusions and heart size;

converting said plurality of clinical parameters and said radiographic descriptors into numerical expressions;

transforming each of said numerical expressions into a number in a predetermined range;

inputting said transformed numerical expressions into a neural network; and

diagnosing at least one of said plurality of interstitial lung diseases using said neural network in accordance with said input expressions.

4. A method as recited in claim 3, further comprising the steps of:

training said neural network to identify each of said plurality of predetermined interstitial lung diseases using said clinical parameters and radiographic descriptors; and

determining an optimal number of hidden layers of said neural network by adjusting said neural network to minimize a difference between an output diagnosis and a desired diagnosis for a given set of said input expressions.

5. A method as recited in claim 4, further comprising the step of:

using receiver operating characteristic (ROC) analysis to determine optimal numbers of hidden layers, learning iterations and units per hidden layer of said neural network.

6. A method as recited in claim 4, wherein said step of training said neural network uses at least 200 learning iterations.

7. A method as recited in claim 3, wherein said neural network includes a plurality of parallel neural networks equal in number to said plurality of predetermined interstitial lung diseases, each of said neural networks having a plurality of inputs and one output, said outputs respectively corresponding to said plurality of predetermined interstitial lung diseases.

8. A method as recited in claim 3, further comprising the steps of:

dividing the diagnosis of said plurality of predetermined interstitial lung diseases into a plurality of stages using a plurality of neural networks.

9. A method as recited in claim 3, wherein said step of selecting a plurality of clinical parameters comprises:

selecting at least plural of the following clinical parameters;

age of a subject,

sex of said subject,

duration of symptoms,

severity of symptoms,

temperature of said subject, and

immune status of said subject.

10. A method as recited in claim 3, wherein said neural network comprises 20 input units, one hidden layer with 6 hidden units and 9 output units, said output units respectively corresponding to 9 interstitial lung diseases.

11. A method as recited in claim 3, further comprising the steps of:

selecting said radiographic descriptors using a computer by computerized automated lung texture analysis; and

interfacing said neural network to said computer.

12. A method for differential diagnosis of plurality of predetermined interstitial lung diseases using an incomplete set of input data, the method comprising the steps of:

selecting a plurality of clinical parameters defining characteristics of a subject;

selecting radiographic descriptors comprising predetermined features obtained from a radiographic chest image defining characteristics of interstitial lung diseases, said radiographic descriptors consisting essentially of each of the following fourteen radiographic descriptors;

distribution of infiltrates in 6 lung zones, homogeneity, fineness, nodularity, septal lines and honeycombing of said infiltrates, and lymphadenopathy, pleural effusions and heart size;

forming a complete set of input data comprising said plurality of clinical parameters and said radiographic descriptors;

converting said plurality of clinical parameters and said radiographic descriptors into numerical expressions;

transforming each of said numerical expressions into a number in a predetermined range;

training a neural network to identify each of said plurality of predetermined interstitial lung diseases using a database of said complete set of input data;

inputting said transformed numerical expressions into said neural network, said input expressions representing incomplete sets of input data; and

diagnosing at least one of said plurality of predetermined interstitial lung diseases in accordance with said input expressions.

13. A method as recited in claim 12, further comprising:

training said neural network to identify each of said plurality of predetermined, diseases using said clinical parameters and radiographic descriptors; and

determining an optimal number of hidden layers of said neural network by adjusting said neural network to minimize a difference between an output diagnosis and a desired diagnosis for a given set of said input expressions.

14. A method as recited in claim 13, further comprising the steps of:

using receiver operating characteristic (ROC) analysis to determine optimal numbers of hidden layers, learning iterations and units per hidden layer of said neural network.

15. A method as recited in claim 13, wherein the step of training said neural network uses at least 200 learning iterations.

16. A method as recited in claim 12, wherein said neural network comprises a plurality of parallel neural networks equal in number to said plurality of predetermined interstitial lung diseases, each of said neural networks having a plurality of inputs and one output, said outputs respectively corresponding to said plurality of predetermined interstitial lung diseases.

17. A method as recited in claim 12, further comprising the step of:

dividing the diagnosis of said plurality of predetermined interstitial lung diseases into a plurality of stages using a plurality of neural networks.

18. A method as recited in claim 12, further comprising the steps of:

completing an incomplete set of input data using a dummy value for each missing piece of input data, said dummy value being in said predetermined range;

inputting said transformed numerical expressions and said dummy values into said neural network, said input expressions representing incomplete sets of input data;

diagnosing at least one of said plurality of predetermined interstitial lung diseases using said neural network in accordance with said input expressions and input dummy values.

19. A method as recited in claim 18, further comprising the steps of:

training said neural network to identify each of said plurality of predetermined interstitial lung diseases using said clinical parameters and said radiographic descriptors; and

determining an optimal number of hidden layers of said neural network by adjusting said neural network to minimize a difference between an output diagnosis and a desired diagnosis for a given set of said input expressions.

20. A method as recited in claim 19, further comprising the step of:

using receiver operating characteristic (ROC) analysis to determine optimal numbers of hidden layers, learning iterations and units per hidden layer of said neural network.

21. A method as recited in claim 19, wherein the step of training said neural network uses at least 200 learning iterations.

22. A method as recited in claim 18, wherein said neural network comprises a plurality of parallel neural networks equal in number to said plurality of predetermined interstitial lung diseases, each of said neural networks having a plurality of inputs and one output, said outputs respectively corresponding to said plurality of predetermined interstitial lung diseases.

23. A method as recited in claim 18, further comprising the step of:

dividing the diagnosis of said plurality of predetermined interstitial lung diseases into a plurality of stages using a plurality of neural networks.

24. A method as recited in claim 18, wherein said neural network comprises 3 layers, 45 input units, at least one hidden unit and at least one output unit.

25. A method for differential diagnosis of a plurality of breast diseases using an incomplete set of input data, the method comprising the steps of:

selecting a plurality of clinical parameters defining characteristics of a subject;

selecting radiographic descriptors comprising predetermined features obtained from a radiographic breast image defining characteristics of breast diseases, said radiographic descriptors consisting essentially of each of the following twelve radiographic descriptors;

shape of density, size of density, margin of density, margin spiculation and pattern of density, presence of, number of, shape of, uniformity of, and distribution of microcalcifications, parenchymal distortion and skin thickening;

forming a complete set of input data comprising said clinical parameters and said plurality of radiographic descriptors;

converting said plurality of clinical parameters and said radiographic descriptors into numerical expressions;

transforming each of said numerical expressions into a number in a predetermined range;

training a neural network to identify each of said plurality of breast diseases using a database of said complete set of input data;

inputting said transformed numerical expressions into a neural network, said input expressions representing incomplete sets of input data; and

diagnosing at least one of said plurality of breast diseases using said neural network in accordance with said input expressions.

26. A method as recited in claim 25 wherein said neural network comprises 3 layers, 45 input units, at least one hidden unit and at least one output unit.