Object detection in paired imaging

1 . A computer readable medium storing instructions for lesion detection in paired medical imaging that when executed by a computing system cause the computing system to:

receive a set of medical images comprising at least two multi-view images of a human body part of a same patient;

detect candidate lesion regions for each medical image of the set, each candidate lesion region having a corresponding initial probability of representing a lesion;

determine a matching probability for each candidate lesion region across the set of images such that candidate lesion regions from one image of the set are assigned matching probabilities with respect to candidate lesion regions of other images of the set;

generate a modified probability for each candidate lesion region according to one or more lesion-specific weighting factors and one or more of the matching probabilities for that candidate lesion region;

adjust the initial probability of each candidate lesion region using the modified probability to generate a refined probability for each candidate lesion region; and

identify the candidate lesion regions in each medical image that have refined probabilities that meet a minimum threshold probability.

2 . The computer readable medium of claim 1 , wherein instructions to determine the matching probability for each candidate lesion region across the set of medical images cause the computing system to:

identify pairs of candidate lesion regions across the set of medical images; and

apply a greedy matching operation to each identified pair.

3 . The computer readable medium of claim 2 , wherein instructions to determine the matching probability for each candidate lesion region across the set of medical images comprise instructions for a Siamese network, the Siamese network being used to identify the pairs across the set of medical images.

4 . The computer readable medium of claim 1 , wherein the one or more lesion-specific weighting factors are predicted by independent regressors with a linear output activation function.

5 . The computer readable medium of claim 1 , wherein instructions to detect the areas of interest for each medical image of the set of medical images comprise instructions implementing a single-view detector.

6 . The computer readable medium of claim 5 , wherein the single-view detector comprises an object detection algorithm for generating an initial detection candidate and a patch classifier for removing false positives.

7 . The computer readable medium of claim 1 , wherein each medical image of the set of medical images is an image of a common target area.

8 . The computer readable medium of claim 7 , wherein the common target area is a breast of a patient, the set of medical images comprising multi-view images of the breast of the patient.

9 . A method of lesion detection in paired medical imaging comprising:

receiving a set of medical images comprising at least two multi-view images of a human body part of a same patient;

detecting candidate lesion regions for each medical image of the set, each candidate lesion region having a corresponding initial probability of representing a lesion;

determining a matching probability for each candidate lesion region across the set of images such that candidate lesion regions from one image of the set are assigned matching probabilities with respect to candidate lesion regions of other images of the set;

generating a modified probability for each candidate lesion region according to one or more lesion-specific weighting factors and one or more of the matching probabilities for that candidate lesion region;

adjusting the initial probability of each candidate lesion region using the modified probability to generate a refined probability for each candidate lesion region; and

identifying the candidate lesion regions in each medical image that have refined probabilities that meet a minimum threshold probability.

10 . The method of claim 9 , wherein determining the matching probability for each candidate lesion region across the set of medical images comprises:

identifying pairs of candidate lesion regions across the set of medical images; and

applying a greedy matching operation to each identified pair.

11 . The method of claim 10 , wherein identifying pairs of candidate lesion regions across the set of medical images comprises utilizing a Siamese network.

12 . The method of claim 9 , wherein the one or more lesion specific weighting factors are predicted by independent regressors with a linear output activation function.

13 . The method of claim 9 , wherein detecting the candidate lesion regions for each image of the set of medical images comprises using a single-view detector.

14 . The method of claim 13 , wherein the single-view detector comprises an object detection algorithm for generating an initial detection candidate and a patch classifier for removing false positives.

15 . The method of claim 9 , wherein each medical image of the set of medical images is an image of a common target area.

16 . The method of claim 15 , wherein the common target area is a breast of a patient, the set of medical images comprising multi-view images of the breast of the patient.

17 . The method of claim 16 , wherein the set of medical images comprises ipsilateral images of the breast of the patient.