Object detection in medical image

Abstract

A device to detect an object in a medical image is described. An image analysis application, executed by the device, receives the medical image as an input. The medical image is next partitioned to sub-regions. Parts of the object are detected in a selection of the sub-regions using a deep-learning neural network (DNN) model. Bounding boxes for the selection are also determined. The bounding boxes are evaluated based on a confidence score detected as above a threshold level. The confidence score designates the parts as contained within the selection. Next, a region of interest (ROI) is determined as a group including the selection. Similar orientations associated with the bounding boxes are comparable to similar orientations of a positive training model of the DNN model. Furthermore, the selection is designated as the ROI within the medical image. The medical image is provided with the ROI to a user.

1 . A method to detect an object in a medical image, the method comprising:
 receiving the medical image as an input;   partitioning the medical image into sub-regions;   detecting parts of an object in a selection of the sub-regions using a deep-learning neural network (DNN) model;   determining bounding boxes for the selection, wherein each of the bounding boxes are evaluated based on a confidence score detected as above a threshold level, and wherein the confidence score designates the parts as contained within the selection;   determining a region of interest (ROI) as a group comprising the selection, wherein similar orientations associated with the bounding boxes are comparable to similar orientations of a positive training model of the DNN model;   designating the selection as the ROI within the medical image; and   providing the medical image with the ROI to a user;   wherein the partitioning and detecting steps are performed before the determining steps, designating step, and providing step.   
     
     
         2 . The method of  claim 1 , wherein the similar orientations associated with the bounding boxes include similar angular orientations between the parts of the object. 
     
     
         3 . The method of  claim 1 , wherein the similar orientations include similar distances between the parts of the object. 
     
     
         4 . The method of  claim 1 , further comprising:
 applying a non-maximum suppression (NMS) mechanism to the selection to obtain a set of object bounding boxes.   
     
     
         5 . The method of  claim 1 , wherein the DNN model associated with a detection of the parts of the object includes a region based convolutional neural network (R-CNN) model, a fast R-CNN model, a faster R-CNN model, a you only look once (YOLO) model, or a single shot multi-box (SSD) model. 
     
     
         6 . The method of  claim 1 , wherein the sub-regions within the ROI are labelled with one or more annotations associated with a type of tissue. 
     
     
         7 . The method of  claim 6 , wherein the type of tissue includes lobulated, spiculated, angular, clear boundary, oval, circumscribed, or abrupt interface. 
     
     
         8 . The method of  claim 6 , wherein the one or more annotations include a relative position associated with a part of the object, and wherein the relative position includes top, bottom, left side, or right side. 
     
     
         9 . The method of  claim 6 , further comprising:
 providing a user interface associated with the ROI, wherein the user interface is configured to allow the user to change the one or more annotations associated with the ROI.   
     
     
         10 . The method of  claim 9 , further comprising:
 detecting the user providing a change to the one or more annotations to personalize the one or more annotations to the user; and   identifying a rate of concordance of the user in relation to the DNN model.   
     
     
         11 . The method of  claim 10 , further comprising:
 determining the concordance rate of the user as above a threshold;   re-training the DNN model based on the change to the one or more annotations;   re-processing the selection based on the change to the one or more annotations; and   re-labelling the sub-regions of the ROI based on the change to the one or more annotations to personalize the one or more annotations to one or more preferences of the user.   
     
     
         12 . The method of  claim 10 , further comprising:
 determining the concordance rate of the user as below a threshold;   rejecting the change to the one or more annotations; and   providing a notification to the user to re-evaluate the change to the one or more annotations.   
     
     
         13 . The method of  claim 1 , wherein the ROI includes a lesion. 
     
     
         14 . The method of  claim 1 , wherein a training mechanism associated with the DNN model includes a compensation for an unbalanced training data consisting of a majority of training medical images with no lesion and a minority of training medical images with a lesion, and wherein the training mechanism includes a down-sampling of the majority, an up-sampling of the minority, or a utilization of a cost sensitive mechanism, a gradient boost machine, or a hard negative mining mechanism. 
     
     
         15 . A device to detect an object in a medical image, wherein the device is configured to:
 receive the medical image as an input;   partition the medical image into sub-regions;   detect parts of the object in a selection of the sub-regions using a deep-learning neural network (DNN) model;   determine bounding boxes for the selection, wherein each of the bounding boxes are evaluated based on a confidence score detected as above a threshold level, and wherein the confidence score designates the parts as contained within the selection;   determine a region of interest (ROI) as a group comprising the selection, wherein similar orientations associated with the bounding boxes are comparable to similar orientations of a positive training model of the DNN model;   designate the selection as the ROI within the medical image;   label the selection within the ROI with one or more annotations associated with a type of tissue, wherein the type of tissue includes lobulated, spiculated, angular, clear boundary, oval, circumscribed, or abrupt interface; and   provide the medical image with the one or more annotations and the ROI to a user;   wherein the partitioning and detecting steps are performed before the determining steps, designating step, labeling step, and providing step.   
     
     
         16 . A device for detecting an object in a medical image, the device comprising:
 a memory configured to store instructions associated with an image analysis application,   a processor coupled to the memory, the processor executing the instructions associated with the image analysis application, wherein the image analysis application includes:
 a computer assisted detection module configured to: 
 receive the medical image as an input; 
 partition the medical image into sub-regions; 
 detect parts of the object in a selection of the sub-regions using a deep-learning neural network (DNN) model; 
 determine bounding boxes for the selection, wherein each of the bounding boxes are evaluated based on a confidence score detected as above a threshold level, and wherein the confidence score designates the parts as contained within the selection; 
 determine a region of interest (ROI) as a group comprising the selection, wherein similar orientations associated with the bounding boxes are comparable to similar orientations of a positive training model of the DNN model; 
 designate the selection as the ROI within the medical image; 
 label the selection within the ROI with one or more annotations associated with a type of tissue, wherein the type of tissue includes lobulated, spiculated, angular, clear boundary, oval, circumscribed, or abrupt interface; and 
 provide the medical image with the one or more annotations and the ROI to a user; 
 wherein the partitioning and detecting steps are performed before the determining steps, designating step, labeling step, and providing step.