Mixed-format labels for pathology detection and localization in magnetic resonance (MR) imaging

1 . A method comprising:

obtaining a radiological image, the radiological image corresponding to an imaged anatomical area;

generating a target map corresponding to a plurality of candidate anatomical defect locations within a cropped radiological image comprising a subset of the radiological image, wherein the target map comprises predicted positional information generated as a prediction output of a semantic segmentation neural network trained based on positional defect data comprising a first label format included in a set of mixed-format labels associated with a plurality of training radiological images;

generating, using a Gaussian kernel of a morphological peak-finding algorithm, a convoluted map including a respective peak for each candidate anatomical defect location of the target map, wherein generating the convoluted map includes convolving the target map with the Gaussian kernel to obtain the respective peak for each candidate anatomical defect location;

identifying, using the morphological peak-finding algorithm and the convoluted map, a particular candidate anatomical defect location from the plurality of candidate anatomical defect locations, wherein the identifying includes selecting, from the convoluted map, a location associated with a greatest peak value in the convoluted map;

generating at least one volume of interest (VOI) centered around the particular candidate anatomical defect location within the cropped radiological image; and

classifying, using a classification neural network trained based on both the positional defect data and categorical defect data comprising a second label format included in the set of mixed-format labels associated with the plurality of training radiological images, the particular candidate anatomical defect location within the cropped radiological image, wherein:

the classification neural network is further trained on a set of VOIs including one or more VOIs each centered around a defect location predicted by the semantic segmentation neural network based on the positional defect data, and one or more VOIs each centered around a ground-truth defect location included in the set of mixed-format labels;

classifying the particular candidate anatomical defect location includes determining categorical information corresponding to a predicted pathology associated with the particular candidate anatomical defect location; and

the categorical information is indicative of a predicted defect type of the predicted pathology and a predicted severity grade for the predicted defect type.

2 . The method of claim 1 , wherein the classification neural network comprises a three-dimensional (3D) ResNet50 convolutional neural network (CNN).

3 . The method of claim 1 , wherein:

the classification neural network is trained using the plurality of training radiological images; and

each respective training radiological image of the plurality of training radiological images is associated with a categorical label comprising the second label format and indicative of an anatomical defect category present in the respective training radiological image.

4 . The method of claim 3 , wherein:

the plurality of training radiological images are associated with categorical labels comprising the second label format and positional labels comprising the first label format of the set of mixed-format labels.

5 . The method of claim 1 , wherein the at least one VOI is generated centered around a candidate anatomical defect coordinate determined based on the target map.

6 . The method of claim 1 , further comprising:

analyzing, using the morphological peak-finding algorithm, each candidate anatomical defect location of the plurality of candidate anatomical defect locations;

identifying, using the morphological peak-finding algorithm, the particular candidate anatomical defect location as a best candidate from the plurality of candidate anatomical defect locations based on having the greatest peak value; and

generating the VOI centered at the predicted positional information predicted by the semantic segmentation neural network for the particular candidate anatomical defect location identified by the morphological peak-finding algorithm as the best candidate.

7 . The method of claim 6 , wherein the classification neural network receives a corresponding VOI centered around each candidate anatomical defect location of the plurality of candidate anatomical defect locations during training.

8 . The method of claim 1 , wherein the semantic segmentation neural network comprises a Residual UNet neural network.

9 . The method of claim 1 , wherein:

the target map is indicative of the plurality of candidate anatomical defect locations; and

each candidate anatomical defect location of the plurality of candidate anatomical defect locations is associated with an isotropic Gaussian sphere.

10 . The method of claim 1 , wherein:

the semantic segmentation neural network is trained using the plurality of training radiological images, wherein each respective training radiological image of the plurality of radiological images is associated with a positional label indicative of a position of an anatomical defect within the respective training radiological image.

11 . The method of claim 10 , wherein the positional label comprises one or more of a point-landmark or a bounding box associated with the position of the anatomical defect within the respective training radiological image.

12 . The method of claim 1 , further comprising:

generating a cropped radiological image based on cropping the radiological image around an anatomical area of interest; and

generating the target map based on processing the cropped radiological image using the semantic segmentation neural network.

13 . The method of claim 12 , wherein cropping the radiological image around the anatomical area of interest comprises:

automatically generating a cropped volume of interest (VOI) centered around the anatomical area of interest, the anatomical area of interest identified in the radiological image using a deep reinforcement learning (DRL) machine learning network.

14 . The method of claim 13 , wherein the cropped VOI includes all of the anatomical area of interest.

15 . The method of claim 12 , wherein the anatomical area of interest comprises an anterior cruciate ligament (ACL) or a medial compartment cartilage (MCC) represented in the radiological image.