Systems and methods for automated detection of visual objects in medical images

A method of training a NN for identification of an indication of at least one visual object in at least one anatomical image of a target individual, comprising:

• providing a training dataset comprising at least one anatomical image for each of a plurality of sample individuals, wherein each at least one anatomical image is associated with per pixel element annotation indicative of a visual object;
• and
• training a NN end-to-end, to perform segmentation by a segmentation component and classification by a classification component according to a size feature computed based on output of the segmentation component and fed into the classification component, and according to a same training dataset fed to both the segmentation component and the classification component, wherein the size feature is computed according to an indication of a respective size of each at least one segmented visual object of a respective training image outputted by the segmentation component,
• wherein the size feature is computed by multiplying each of a plurality of per pixel element segmentation probabilities outputted by the segmentation component, by a volume associated with each pixel element to compute a plurality of sub-volumes, and summing the plurality of sub-volumes, wherein each segmentation probability is indicative of a respective probability of the respective pixel element indicative of the at least one visual object.

A system for identification of an indication of at least one visual object in at least one anatomical image of a target individual, comprising:

• a non-transitory memory having stored thereon a code for execution by at least one hardware processor, the code comprising:
• code for receiving at least one anatomical image of a body portion of a target individual;
• code for inputting the at least one anatomical image into a classification component of a neural network (NN) and into a segmentation component of the NN;
• code for feeding a size feature into the classification component of the NN, wherein the size feature comprises an indication of a respective size of each at least one segmented visual object identified in the at least one anatomical image, the size feature computed according to segmentation data outputted by the segmentation component for each pixel element of the at least one anatomical image;
• code for training a NN for identification of an indication of at least one visual object in at least one anatomical image of a target individual, comprising:
• code for accessing a training dataset comprising at least one anatomical image for each of a plurality of sample individuals, wherein each at least one anatomical image is associated with per pixel element annotation indicative of a visual object;
• and
• code for training a NN end-to-end, to perform segmentation by a segmentation component and classification by a classification component according to a size feature computed based on output of the segmentation component and fed into the classification component, and according to a same training dataset fed to both the segmentation component and the classification component, wherein the size feature is computed according to an indication of a respective size of each at least one segmented visual object of a respective training image outputted by the segmentation component;
• and
• code for computing, by the classification component of the NN, an indication of at least one visual object in the at least one anatomical image
• wherein the size feature is computed by multiplying each of a plurality of per pixel element segmentation probabilities outputted by the segmentation component, by a volume associated with each pixel element to compute a plurality of sub-volumes, and summing the plurality of sub-volumes, wherein each segmentation probability is indicative of a respective probability of the respective pixel element indicative of the at least one visual object.