Granted Patent B2

US 12,444,049 · App. 18/346,662 · Granted Oct 14, 2025

Inventors: Manu Sheoran (Noida, IN); Monika Sharma (Noida, IN); Lovekesh Vig (Noida, IN)

Assignee: TATA CONSULTANCY SERVICES LIMITED

G06T7/0012G06T7/60G06T7/73G06V10/764G06V10/7715G06V10/778G06V10/82G06V20/70G16H30/40G16H50/20G06T2207/10072G06T2207/20081G06T2207/20084G06T2207/30096G06V2201/03

Abstract

Claims (120)

1. A processor implemented method, the method comprising:

receiving, by one or more hardware processors, an input image, wherein the input image is obtained from a diagnostic medical imaging equipment;

extracting, by the one or more hardware processors, a plurality of multi-scale feature maps from the input image using a Feature Pyramid Network (FPN) based feature extraction framework;

generating, by the one or more hardware processors, a classification map based on the plurality of multi-scale feature maps using a Fully Connected Classifier Network (FCCN), wherein the FCCN classifies each of a plurality of pixels associated with each of the plurality of multi-scale feature maps into one of, a) a foreground pixel, and b) a background pixel based on a corresponding conditional probability;

computing, by the one or more hardware processors, a 4D vector corresponding to each of a plurality of foreground pixels using a bounding box regressor network, wherein the 4D vector encodes a location of a corresponding bounding box;

predicting, by the one or more hardware processors, an objectness score corresponding to each of the plurality of foreground pixels using a Fully Connected Prediction Network (FCPN), wherein the objectness score is a confidence score for being one of, the foreground pixel and the background pixel;

computing, by the one or more hardware processors, a centerness score for each of the plurality of foreground pixels using a single centerness network, wherein the centerness score represents a distance between the pixel and a center of a corresponding ground truth bounding box;

computing, by the one or more hardware processors, an updated objectness score for each of the plurality of foreground pixels by multiplying a corresponding centerness score with the corresponding predicted objectness score; and

detecting, by the one or more hardware processors, a plurality of multi-sized lesions in the input image based on the updated objectness score corresponding to each of the plurality of foreground pixels and corresponding 4D vector using a trained few-shot adversarial lesion detector network, wherein the few-shot adversarial lesion detector network is trained using a periodic gradient updation based overfitting aware few-shot learning mechanism.

2. The processor implemented method of claim 1 , wherein the few-shot adversarial lesion detector network is trained by:

receiving a labelled dataset pertaining to a source domain, an unlabelled dataset pertaining to a target domain, and a labelled dataset pertaining to the target domain, wherein the source domain and the target domain are distinct and related;

extracting a plurality of multi-scale source feature maps from the labelled dataset pertaining to the source domain using the Feature Pyramid Network (FPN);

extracting a first plurality of multi-scale target feature maps from the unlabelled dataset pertaining to the target domain using the Feature Pyramid Network (FPN);

extracting a second plurality of multi-scale target feature maps from the labelled dataset pertaining to the target domain using the Feature Pyramid Network (FPN);

generating a classification map based on the plurality of multi-scale source feature maps and the second plurality of multi-scale target feature maps using the FCCN, wherein the FCCN classifies each of a plurality of pixels associated with each of the plurality of multi-scale source feature maps and second plurality of multi-scale target feature maps into one of, a) the foreground pixel, and b) the background pixel based on the corresponding conditional probability;

generating the 4D vector corresponding to each of a plurality of foreground pixels using the bounding box regressor network;

computing an updated objectness score for each of the plurality of foreground pixels by multiplying a corresponding centerness score with a corresponding predicted objectness score, wherein the objectness score is predicted based on the plurality of foreground pixels using the fully connected layer and, wherein the centerness score is computed for each of the plurality of foreground pixels using the single centerness network;

training an adversarial lesion detector network based on the 4D vector and updated objectness score in a few-shot manner until a minimum supervised loss is obtained;

training a global discriminator network based on the plurality of multi-scale source feature maps and the first plurality of multi-scale target feature maps until a minimum domain prediction loss is obtained, wherein the global discriminator network is a convolutional neural network comprising a gradient reversal layer;

generating a centre-aware map by multiplying a pixelwise objectness map and an objectness score based centerness map;

training a center-aware discriminator network based on the centre-aware map until a minimum domain prediction loss is obtained, wherein the center-aware discriminator network is the convolutional neural network comprising the gradient reversal layer;

updating the adversarial lesion detector network based on a corresponding domain predictions losses associated with the global discriminator network and the center-aware discriminator network; and

obtaining the few-shot adversarial lesion detector network by performing a supervised learning on the updated adversarial lesion detector network based on the plurality of multi-scale source feature maps and the second plurality of multi-scale target feature maps using a periodic gradient updation based overfitting aware few-shot learning mechanism.

3. The processor implemented method of claim 2 , wherein obtaining the few-shot adversarial lesion detector network by performing the supervised learning on the updated adversarial lesion detector network based on the plurality of multi-scale source feature maps and the second plurality of multi-scale target feature maps using a periodic gradient updation based overfitting aware few-shot learning mechanism comprises:

repeatedly performing for a predefined plurality of iterations:

initializing an iteration value, wherein the iteration value is updated after each iteration of the plurality of iterations;

training the adversarial lesion detector network based on the plurality of multi-scale source feature maps;

updating a plurality of hyperparameters associated with the adversarial lesion detector network;

computing a decision value based on a plurality of deciding parameters, wherein the plurality of deciding parameters comprises a total number of training samples from the source domain, a total number of few-shot training samples from the target domain and the plurality of hyperparameters;

training the adversarial lesion detector network based on the plurality of multi-scale target feature maps if the iteration value is completely divisible by the decision value, wherein iteration value updation and training of the adversarial lesion detector network is performed if the iteration value is not completely divisible by the decision value; and

updating the plurality of hyperparameters associated with the adversarial lesion detector network.

4. The processor implemented method of claim 2 , wherein generating the centre-aware map based on the pixelwise objectness map and the objectness score based centerness map comprises:

predicting the pixelwise objectness map based on the plurality of multi-scale source feature maps and the first plurality of multi-scale target feature maps using the adversarial lesion detector network;

predicting an objectness score based centerness map based on the plurality of multi-scale source feature maps and the first plurality of multi-scale target feature maps using the adversarial lesion detector network; and

generating the centre-aware map by multiplying the pixelwise objectness map and the objectness score based centerness map.

5. The processor implemented method of claim 1 , further comprising a pseudo label generation by:

generating a plurality of pseudo-labels based on the unlabelled dataset pertaining to the target domain using few-shot adversarial lesion detector network, wherein each of the plurality of pseudo-labels comprises a corresponding 4D vector and an objectness score;

selecting a plurality of confident pseudo-labels from the plurality of pseudo-labels based on a predefined confidence threshold, wherein the plurality of pseudo-labels with the corresponding abjectness score greater than the predefined confidence threshold are selected; and

updating the labelled dataset pertaining to the target domain with the plurality of confident pseudo-labels, wherein the few-shot adversarial lesion detector network is further trained using the updated labelled dataset pertaining to the target domain.

6. A system comprising:

at least one memory storing programmed instructions;

one or more Input/Output (I/O) interfaces; and

one or more hardware processors operatively coupled to the at least one memory, wherein the one or more hardware processors are configured by the programmed instructions to:

receive an input image, wherein the input age is obtained from a diagnostic medical imaging equipment;

extract a plurality of multi-scale feature maps from the input image using a Feature Pyramid Network (FPN) based feature extraction framework;

generate a classification map based on the plurality of multi-scale feature maps using a Fully Connected Classifier Network (FCCN), wherein the FCCN classifies each of a plurality of pixels associated with each of the plurality of multi-scale feature maps into one of, a) a foreground pixel, and b) a background pixel based on a corresponding conditional probability;

compute a 4D vector corresponding to each of a plurality of foreground pixels using a bounding box regressor network, wherein the 4D vector encodes a location of a corresponding bounding box;

predict an objectness score corresponding to each of the plurality of foreground pixels using a Fully Connected Prediction Network (FCPN), wherein the objectness score is a confidence score for being one of, the foreground pixel and the background pixel;

compute a centerness score for each of the plurality of foreground pixels using a single centerness network, wherein the centerness score represents a distance between the pixel and a center of a corresponding ground truth bounding box;

compute an updated objectness score for each of the plurality of foreground pixels by multiplying a corresponding centerness score with the corresponding predicted objectness score; and

detect a plurality of multi-sized lesions in the input image based on the updated objectness score corresponding to each of the plurality of foreground pixels and corresponding 4D vector using a trained few-shot adversarial lesion detector network, wherein the few-shot adversarial lesion detector network is trained using a periodic gradient updation based overfitting aware few-shot learning mechanism.

7. The system of claim 6 , wherein the few-shot adversarial lesion detector network is trained by:

receiving a labelled dataset pertaining to a source domain, an unlabelled dataset pertaining to a target domain, and a labelled dataset pertaining to the target domain, wherein the source domain and the target domain are distinct and related;

extracting a plurality of multi-scale source feature maps from the labelled dataset pertaining to the source domain using the Feature Pyramid Network (FPN);

extracting a first plurality of multi-scale target feature maps from the unlabelled dataset pertaining to the target domain using the Feature Pyramid Network (FPN);

extracting a second plurality of multi-scale target feature maps from the labelled dataset pertaining to the target domain using the Feature Pyramid Network (FPN);

generating a classification map based on the plurality of multi-scale source feature maps and the second plurality of multi-scale target feature maps using the FCCN, wherein the FCCN classifies each of a plurality of pixels associated with each of the plurality of multi-scale source feature maps and second plurality of multi-scale target feature maps into one of, a) the foreground pixel, and b) the background pixel based on the corresponding conditional probability;

generating the 4D vector corresponding to each of a plurality of foreground pixels using the bounding box regressor network;

computing an updated objectness score for each of the plurality of foreground pixels by multiplying a corresponding centerness score with a corresponding predicted objectness score, wherein the objectness score is predicted based on the plurality of foreground pixels using the fully connected layer and, wherein the centerness score is computed for each of the plurality of foreground pixels using the single centerness network;

training an adversarial lesion detector network based on the 4D vector and updated objectness score in a few-shot manner until a minimum supervised loss is obtained;

training a global discriminator network based on the plurality of multi-scale source feature maps and the first plurality of multi-scale target feature maps until a minimum domain prediction loss is obtained, wherein the global discriminator network is a convolutional neural network comprising a gradient reversal layer;

generating a centre-aware map by multiplying a pixelwise objectness map and an objectness score based centerness map;

training a center-aware discriminator network based on the centre-aware map until a minimum domain prediction loss is obtained, wherein the center-aware discriminator network is the convolutional neural network comprising the gradient reversal layer;

updating the adversarial lesion detector network based on a corresponding domain predictions losses associated with the global discriminator network and the center-aware discriminator network; and

obtaining the few-shot adversarial lesion detector network by performing a supervised learning on the updated adversarial lesion detector network based on the plurality of multi-scale source feature maps and the second plurality of multi-scale target feature maps using a periodic gradient updation based overfitting aware few-shot learning mechanism.

8. The system of claim 7 , wherein obtaining the few-shot adversarial lesion detector network by performing the supervised learning on the updated adversarial lesion detector network based on the plurality of multi-scale source feature maps and the second plurality of multi-scale target feature maps using a periodic gradient updation based overfitting aware few-shot learning mechanism comprises:

repeatedly performing for a predefined plurality of iterations:

initializing an iteration value, wherein the iteration value is updated after each iteration of the plurality of iterations;

training the adversarial lesion detector network based on the plurality of multi-scale source feature maps;

updating a plurality of hyperparameters associated with the adversarial lesion detector network;

computing a decision value based on a plurality of deciding parameters, wherein the plurality of deciding parameters comprises a total number of training samples from the source domain, a total number of few-shot training samples from the target domain and the plurality of hyperparameters;

training the adversarial lesion detector network based on the plurality of multi-scale target feature maps if the iteration value is completely divisible by the decision value, wherein iteration value updation and training of the adversarial lesion detector network is performed if the iteration value is not completely divisible by the decision value; and

updating the plurality of hyperparameters associated with the adversarial lesion detector network.

9. The system of claim 7 , wherein generating the centre-aware map based on the pixelwise objectness map and the objectness score based centerness map comprises:

predicting the pixelwise objectness map based on the plurality of multi-scale source feature maps and the first plurality of multi-scale target feature maps using the adversarial lesion detector network;

predicting an objectness score based centerness map based on the plurality of multi-scale source feature maps and the first plurality of multi-scale target feature maps using the adversarial lesion detector network; and

generating the centre-aware map by multiplying the pixelwise objectness map and the objectness score based centerness map.

10. The system of claim 6 , further comprising a pseudo label generation by:

generating a plurality of pseudo-labels based on the unlabelled dataset pertaining to the target domain using few-shot adversarial lesion detector network, wherein each of the plurality of pseudo-labels comprises a corresponding 4D vector and an objectness score;

selecting a plurality of confident pseudo-labels from the plurality of pseudo-labels based on a predefined confidence threshold, wherein the plurality of pseudo-labels with the corresponding abjectness score greater than the predefined confidence threshold are selected; and

updating the labelled dataset pertaining to the target domain with the plurality of confident pseudo-labels, wherein the few-shot adversarial lesion detector network is further trained using the updated labelled dataset pertaining to the target domain.

11. One or more non-transitory machine readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors causes:

receiving an input image, wherein the input image is obtained from a diagnostic medical imaging equipment;

extracting a plurality of multi-scale feature maps from the input image using a Feature Pyramid Network (FPN) based feature extraction framework;

generating a classification map based on the plurality of multi-scale feature maps using a Fully Connected Classifier Network (FCCN), wherein the FCCN classifies each of a plurality of pixels associated with each of the plurality of multi-scale feature maps into one of, a) a foreground pixel, and b) a background pixel based on a corresponding conditional probability;

computing, by the one or more hardware processors, a 4D vector corresponding to each of a plurality of foreground pixels using a bounding box regressor network, wherein the 4D vector encodes a location of a corresponding bounding box;

predicting an objectness score corresponding to each of the plurality of foreground pixels using a Fully Connected Prediction Network (FCPN), wherein the objectness score is a confidence score for being one of, the foreground pixel and the background pixel;

computing a centerness score for each of the plurality of foreground pixels using a single centerness network, wherein the centerness score represents a distance between the pixel and a center of a corresponding ground truth bounding box;

computing an updated objectness score for each of the plurality of foreground pixels by multiplying a corresponding centerness score with the corresponding predicted objectness score; and

detecting a plurality of multi-sized lesions in the input image based on the updated objectness score corresponding to each of the plurality of foreground pixels and corresponding 4D vector using a trained few-shot adversarial lesion detector network, wherein the few-shot adversarial lesion detector network is trained using a periodic gradient updation based overfitting aware few-shot learning mechanism.

12. The one or more non-transitory machine readable information storage mediums of claim 11 , wherein the few-shot adversarial lesion detector network is trained by:

receiving a labelled dataset pertaining to a source domain, an unlabelled dataset pertaining to a target domain, and a labelled dataset pertaining to the target domain, wherein the source domain and the target domain are distinct and related;

extracting a plurality of multi-scale source feature maps from the labelled dataset pertaining to the source domain using the Feature Pyramid Network (FPN);

extracting a first plurality of multi-scale target feature maps from the unlabelled dataset pertaining to the target domain using the Feature Pyramid Network (FPN);

extracting a second plurality of multi-scale target feature maps from the labelled dataset pertaining to the target domain using the Feature Pyramid Network (FPN);

generating a classification map based on the plurality of multi-scale source feature maps and the second plurality of multi-scale target feature maps using the FCCN, wherein the FCCN classifies each of a plurality of pixels associated with each of the plurality of multi-scale source feature maps and second plurality of multi-scale target feature maps into one of, a) the foreground pixel, and b) the background pixel based on the corresponding conditional probability;

generating the 4D vector corresponding to each of a plurality of foreground pixels using the bounding box regressor network;

computing an updated objectness score for each of the plurality of foreground pixels by multiplying a corresponding centerness score with a corresponding predicted objectness score, wherein the objectness score is predicted based on the plurality of foreground pixels using the fully connected layer and, wherein the centerness score is computed for each of the plurality of foreground pixels using the single centerness network;

training an adversarial lesion detector network based on the 4D vector and updated objectness score in a few-shot manner until a minimum supervised loss is obtained;

training a global discriminator network based on the plurality of multi-scale source feature maps and the first plurality of multi-scale target feature maps until a minimum domain prediction loss is obtained, wherein the global discriminator network is a convolutional neural network comprising a gradient reversal layer;

generating a centre-aware map by multiplying a pixelwise objectness map and an objectness score based centerness map;

training a center-aware discriminator network based on the centre-aware map until a minimum domain prediction loss is obtained, wherein the center-aware discriminator network is the convolutional neural network comprising the gradient reversal layer;

updating the adversarial lesion detector network based on a corresponding domain predictions losses associated with the global discriminator network and the center-aware discriminator network; and

obtaining the few-shot adversarial lesion detector network by performing a supervised learning on the updated adversarial lesion detector network based on the plurality of multi-scale source feature maps and the second plurality of multi-scale target feature maps using a periodic gradient updation based overfitting aware few-shot learning mechanism.

13. The one or more non-transitory machine readable information storage mediums of claim 12 , wherein obtaining the few-shot adversarial lesion detector network by performing the supervised learning on the updated adversarial lesion detector network based on the plurality of multi-scale source feature maps and the second plurality of multi-scale target feature maps using a periodic gradient updation based overfitting aware few-shot learning mechanism comprises:

repeatedly performing for a predefined plurality of iterations:

initializing an iteration value, wherein the iteration value is updated after each iteration of the plurality of iterations;

training the adversarial lesion detector network based on the plurality of multi-scale source feature maps;

updating a plurality of hyperparameters associated with the adversarial lesion detector network;

computing a decision value based on a plurality of deciding parameters, wherein the plurality of deciding parameters comprises a total number of training samples from the source domain, a total number of few-shot training samples from the target domain and the plurality of hyperparameters;

training the adversarial lesion detector network based on the plurality of multi-scale target feature maps if the iteration value is completely divisible by the decision value, wherein iteration value updation and training of the adversarial lesion detector network is performed if the iteration value is not completely divisible by the decision value; and

updating the plurality of hyperparameters associated with the adversarial lesion detector network.

14. The one or more non-transitory machine readable information storage mediums of claim 12 , wherein generating the centre-aware map based on the pixelwise abjectness map and the abjectness score based centerness map comprises:

predicting the pixelwise abjectness map based on the plurality of multi-scale source feature maps and the first plurality of multi-scale target feature maps using the adversarial lesion detector network;

predicting an abjectness score based centerness map based on the plurality of multi-scale source feature maps and the first plurality of multi-scale target feature maps using the adversarial lesion detector network; and

generating the centre-aware map by multiplying the pixelwise abjectness map and the abjectness score based centerness map.

15. The one or more non-transitory machine readable information storage mediums of claim 11 , further comprising a pseudo label generation by:

generating a plurality of pseudo-labels based on the unlabelled dataset pertaining to the target domain using few-shot adversarial lesion detector network, wherein each of the plurality of pseudo-labels comprises a corresponding 4D vector and an abjectness score;

selecting a plurality of confident pseudo-labels from the plurality of pseudo-labels based on a predefined confidence threshold, wherein the plurality of pseudo-labels with the corresponding abjectness score greater than the predefined confidence threshold are selected; and

updating the labelled dataset pertaining to the target domain with the plurality of confident pseudo-labels, wherein the few-shot adversarial lesion detector network is further trained using the updated labelled dataset pertaining to the target domain.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (13)