Granted Patent B2

US 12,493,667 · App. 17/039,063 · Granted Dec 9, 2025

Inventor: Aviad Zlotnick (Mitzpeh Netofah, IL)

Assignee: International Business Machines Corporation

G06F18/2433G06F17/18G06F18/211G06F18/214G06N3/04G06N3/08

Abstract

Claims (98)

1 . A computer-implemented method for outlier detection, comprising:

receiving, by one or more processors, an input item comprising an X-ray image;

running, by the one or more processors, a prediction on a training dataset including the input item by utilizing a deep neural network model;

performing, by the one or more processors, a learning process on a mapping function from inputs to outputs by updating weights of a network in response to errors that occur in the prediction by the deep neural network model on the training dataset;

continually updating, by the one or more processor, the weights of the network to reduce the errors to a predetermined level;

extracting, by the one or more processors, a first feature vector from the input item;

quantizing, by the one or more processors, the extracted first feature vector to discrete values, the discrete values being below or above a median value of the extracted first feature vector from the input item;

computing, by the one or more processors, a first number of special t-way feature combinations in the input item;

computing, by the one or more processors, a computed threshold comprising calculating a mean (m) of a count of rare t-way feature combinations for every input item in a training dataset plus a standard deviation(s) of the count of rare t-way feature combinations plus a square root (sqrt) of the count of rare t-way feature combinations per input item in the training dataset, wherein the count of rare t-way feature combinations does not include the occurrence of feature combinations in the input item;

comparing, by the one or more processors, the computed first number of special t-way feature combinations against the computed threshold;

based on the comparison, flagging, by the one or more processors, the input item as an outlier comprising an abnormal X-ray image with an impact on normal input items being reduced by at least one predetermined tuning threshold;

filtering, by the one or more processors, the input item as the outlier comprising the abnormal X-ray image from the training dataset;

training, by the one or more processors, the deep neural network model based on the training dataset with the filtered input item; and

generating, by the one or more processors, an alert comprising the flagged input item, wherein the alert comprises a video notification.

2 . The method of claim 1 , wherein the computed threshold is computed as the mean (m) of the count of a number of significant t-way feature combinations minus a multiple of the standard deviations(s), and

wherein running the deep neural network on the input item further comprises:

training the deep neural network for outlier detection, the training comprising:

running, by the one or more processors, the deep neural network on a training dataset;

extracting, by the one or more processors, a second feature vector from the training dataset;

quantizing, by the one or more processors, the extracted second feature vector to discrete values;

determining, by the one or more processors, a count of occurrences of t-way feature combinations in the training dataset, wherein the count of occurrences of t-way feature combinations in the training dataset does not include the occurrence of the feature combination in the training dataset;

based on the count of occurrences of the t-way feature combinations in the training dataset, defining, by the one or more processors, a second number of special t-way feature combinations;

calculating, by the one or more processors, the computed threshold for the defined second number of special t-way feature combinations per training input item in the training dataset; and

tuning, by the one or more processors, the at least one predetermined tuning threshold to reduce flagging of the normal input items.

3 . The method of claim 2 , wherein the first number of special t-way feature combinations in the input item and the second number of special t-way feature combinations in the training dataset comprise a number of rare t-way feature combinations that occur less than n times, for every n<N, wherein N represents a number of t-way combinations of five or fewer.

4 . The method of claim 3 , wherein the computed threshold for the second number of special t-way feature combinations equals the computed threshold for the first number of special t-way feature combinations.

5 . The method of claim 4 , wherein comparing the computed first number of special t-way feature combinations against the computed threshold further comprises:

based on the count of rare t-way feature combinations exceeding the computed threshold, flagging, by the one or more processors, the input item as outlier.

6 . The method of claim 2 , wherein the first number of special t-way feature combinations in the input item and the second number of special t-way feature combinations in the training dataset comprise a number of significant t-way feature combinations, the number of significant t-way feature combinations comprising t-way combinations of features that appear together more than it is suggested by one or more probabilities.

7 . The method of claim 6 , wherein calculating the computed threshold for the second number of special t-way feature combinations in the training dataset comprises:

calculating, by the one or more processors, a mean (m) of a count of the number of significant t-way feature combinations per input item in the training dataset minus five times a standard deviation(s) of the count of the number of significant t-way feature combinations wherein the count of the number of significant t-way feature combinations does not include the occurrence of the feature combinations in the input item.

8 . The method of claim 7 , wherein comparing the computed first number of special t-way feature combinations against the computed threshold further comprises:

based on the count of significant t-way feature combinations in the input item being lower than the computed threshold, flagging, by the one or more processors, the input item as an outlier

filtering input items that are outliers with respect to a majority of the input items based on the flagging;

tuning the computed threshold based on the filtering; and

based on the filtering, flagging, by the one or more processors, the input item as an outlier.

9 . The method of claim 1 , wherein the training dataset comprises X-ray images, the alert comprises haptic feedback, and the predetermined level corresponds with a level in which the learning process gets stuck.

10 . A computer system for outlier detection, comprising:

one or more processors, one or more computer-readable memories, one or more computer-readable tangible storage devices, and program instructions stored on at least one of the one or more storage devices for execution by at least one of the one or more processors via at least one of the one or more memories, wherein the computer system is capable of performing a method comprising:

receiving, by one or more processors, an input item comprising an X-ray image;

running, by the one or more processors, a prediction on a training dataset including the input item by utilizing a deep neural network model;

performing, by the one or more processors, a learning process on a mapping function from inputs to outputs by updating weights of a network in response to errors that occur in the prediction by the deep neural network model on the training dataset;

continually updating, by the one or more processor, the weights of the network to reduce the errors to a predetermined level;

extracting, by the one or more processors, a first feature vector from the input item;

quantizing, by the one or more processors, the extracted first feature vector to discrete values, the discrete values being below or above a median value of the extracted first feature vector from the input item;

computing, by the one or more processors, a first number of special t-way feature combinations in the input item;

computing, by the one or more processors, a computed threshold comprising calculating a mean (m) of a count of rare t-way feature combinations for every input item in a training dataset plus a standard deviation(s) of the count of rare t-way feature combinations plus a square root (sqrt) of the count of rare t-way feature combinations per input item in the training dataset, wherein the count does not include the occurrence of feature combinations in the input item;

comparing, by the one or more processors, the computed first number of special t-way feature combinations against the computed threshold;

based on the comparison, flagging, by the one or more processors, the input item as an outlier comprising an abnormal X-ray image with an impact on normal input items being reduced by at least one predetermined tuning threshold;

filtering, by the one or more processors, the input item as the outlier comprising the abnormal X-ray image from the training dataset;

training, by the one or more processors, the deep neural network model based on the training dataset with the filtered input item; and

generating, by the one or more processors, an alert comprising the flagged input item, wherein the alert comprises an audio notification.

11 . The computer system of claim 10 , wherein the audio notification comprises audible information regarding the flagged input item, and

wherein running the deep neural network on the input item further comprises:

training the deep neural network for outlier detection, the training comprising:

running, by the one or more processors, the deep neural network on a training dataset;

extracting, by the one or more processors, a second feature vector from the training dataset;

quantizing, by the one or more processors, the extracted second feature vector to discrete values;

determining, by the one or more processors, a count of occurrences of t-way feature combinations in the training dataset;

based on the count of occurrences of the t-way feature combinations in the training dataset, defining, by the one or more processors, a second number of special t-way feature combinations;

calculating, by the one or more processors, the computed threshold for the defined second number of special t-way feature combinations per training input item in the training dataset; and

tuning, by the one or more processors, the at least one predetermined tuning threshold to reduce flagging of the normal input items.

12 . The computer system of claim 11 , wherein the first number of special t-way feature combinations in the input item and the second number of special t-way feature combinations in the training dataset comprise a number of rare t-way feature combinations that occur less than n times, for every n<N.

13 . The computer system of claim 12 , wherein calculating the computed threshold for the second number of special t-way feature combinations in the training dataset comprises:

calculating, by the one or more processors, a mean (m) of a count of rare t-way feature combinations for every input item in the training dataset plus a standard deviation(s) of the count of rare t-way feature combinations plus a square root (sqrt) of the count of rare t-way feature combinations per input item in the training dataset.

14 . The computer system of claim 13 , wherein comparing the computed first number of special t-way feature combinations against the computed threshold further comprises:

based on the count of rare t-way feature combinations exceeding the computed threshold, flagging, by the one or more processors, the input item as an outlier.

15 . The computer system of claim 11 , wherein the first number of special t-way feature combinations in the input item and the second number of special t-way feature combinations in the training dataset comprise a number of significant t-way feature combinations, the number of significant t-way feature combinations comprising t-way combinations of features that appear together more than it is suggested by one or more probabilities.

16 . The computer system of claim 15 , wherein calculating the computed threshold for the second number of special t-way feature combinations in the training dataset comprises:

calculating, by the one or more processors, a mean (m) of a count of the number of significant t-way feature combinations per input item in the training dataset minus five times a standard deviation(s) of the count of the number of significant t-way feature combinations.

17 . The computer system of claim 16 , wherein comparing the computed first number of special t-way feature combinations against the computed threshold further comprises:

based on the count of significant t-way feature combinations in the input item being lower than the computed threshold, flagging, by the one or more processors, the input item as an outlier.

18 . The computer system of claim 10 , wherein the training dataset comprises X-ray images, the alert comprises haptic feedback, and the predetermined level corresponds with a level in which the learning process gets stuck.

19 . A computer program product for outlier detection, comprising:

one or more computer readable storage media, and program instructions collectively stored on the one or more computer readable storage media, the program instructions comprising:

program instructions to receive, by one or more processors, an input item comprising an X-ray image;

program instructions to run, by the one or more processors, a prediction on a training dataset including the input item by utilizing a deep neural network model;

program instructions to perform, by the one or more processors, a learning process on a mapping function from inputs to outputs by updating weights of a network in response to errors that occur in the prediction by the deep neural network model on the training dataset;

program instructions to continually update, by the one or more processor, the weights of the network to reduce the errors to a predetermined level;

program instructions to extract, by the one or more processors, a first feature vector from the input item;

program instructions to quantize, by the one or more processors, the extracted first feature vector to discrete values, the discrete values being below or above a median value of the extracted first feature vector from the input item;

program instructions to compute, by the one or more processors, a first number of special t-way feature combinations in the input item;

computing, by the one or more processors, a computed threshold based on a function of a mean (m) of a count of rare t-way feature combinations for every input item in a training dataset and a standard deviation(s) of the count of rare t-way feature combinations, wherein the count does not include the occurrence of feature combinations in the input item;

program instructions to compare, by the one or more processors, the computed first number of special t-way feature combinations against the computed threshold;

based on the comparison, program instructions to flag, by the one or more processors, the input item as an outlier comprising an abnormal X-ray image with an impact on normal input items being reduced by at least one predetermined tuning threshold;

program instructions to filter, by the one or more processors, the input item as the outlier comprising the abnormal X-ray image from the training dataset;

program instructions to train, by the one or more processors, the deep neural network model based on the training dataset with the filtered input item; and

program instructions to generate, by the one or more processors, an alert comprising the flagged input item, wherein the alert comprises a video notification and an audio notification.

20 . The computer program product of claim 19 , wherein the program instructions to run the deep neural network on the input item further comprises:

program instructions to train the deep neural network for outlier detection comprising:

program instructions to run, by the one or more processors, the deep neural network on a training dataset;

program instructions to extract, by the one or more processors, a second feature vector from the training dataset;

program instructions to quantize, by the one or more processors, the extracted second feature vector to discrete values;

program instructions to determine, by the one or more processors, a count of occurrences of t-way feature combinations in the training dataset;

based on the count of occurrences of the t-way feature combinations in the training dataset, program instructions to define, by the one or more processors, a second number of special t-way feature combinations;

program instructions to calculate, by the one or more processors, the computed threshold for the defined second number of special t-way feature combinations per training input item in the training dataset; and

program instructions to tune, by the one or more processors, the at least one predetermined tuning threshold to reduce flagging of the normal input items,

wherein the predetermined level corresponds with a level in which the learning process gets stuck.

Assignments (1)

Continuity (1)

References Cited (38)