System for detecting micro-neuromas and methods of use thereof

A method of (i) assessing a treatment regimen for neuropathic corneal pain, or (ii) for determining the efficacy of a treatment for neuropathic corneal pain, wherein the method of part (i) comprises:

• a) acquiring a first set of in vivo confocal microscopy images of an ocular surface of a subject experiencing neuropathic corneal pain;
• b) analyzing the first set of in vivo confocal microscopy images using a computer programmed with a neural network to identify an anatomical structure associated with neuropathic corneal pain;
• c) administering a therapeutic agent suitable for treating neuropathic corneal pain to the subject for a therapeutically sufficient duration;
• d) acquiring a second set of in vivo confocal microscopy images of the ocular surface of the subject;
• e) analyzing the second set of in vivo confocal microscopy images using a computer programmed with a neural network to determine structural changes in the anatomical structure associated with neuropathic corneal pain;
• and
• f) repeating steps a-e until the subject experiences a reduction in neuropathic corneal pain, wherein the reduction of neuropathic pain is caused by a reduction in at least one dimension of the anatomical structure identified by the in vivo confocal microscopy imaging, and the method of part (ii) comprises:
• a) diagnosing a subject as having or potentially having neuropathic corneal pain;
• b) acquiring a first set of in vivo confocal microscopy images of an ocular surface of the subject;
• c) analyzing the first set of in vivo confocal microscopy images using a computer programmed with a neural network to identify at least one parameter associated with neuropathic corneal pain;
• d) administering a therapeutic agent to the subject;
• e) acquiring a second set of in vivo confocal microscopy images of the ocular surface of the subject;
• f) analyzing the second set of in vivo confocal microscopy images using a computer programmed with a neural network to determine a change in the at least one parameter associated with neuropathic corneal pain;
• and
• g) storing or providing an output indicative of the efficacy of the treatment for neuropathic corneal pain.

A non-transitory computer readable medium having instructions stored thereon, wherein

• (I) the instructions, when executed by a processor, perform a method for automatically determining the presence of at least one neuroma on at least one image of an ocular surface of a subject, the method comprising:
• a) acquiring at least one image of an ocular surface of a subject;
• and
• b) determining the presence of a neuroma on the at least one image of an ocular surface of a subject by analyzing the at least one image of an ocular surface of a subject using a trained neural network, wherein the trained neural network comprises:
• i) a residual learning architecture;
• and
• ii) a backpropagation algorithm comprising a gradient descent optimizer,
• wherein the neural network is trained using a plurality of images of an ocular surface from a population of subjects, wherein the plurality of images of an ocular surface from a population of subjects are augmented using data blending augmentation such as mix-up or data interpolating augmentation prior to training the neural network;
• or
• (II) the non-transitory computer readable medium comprises:
• a) a neural network comprising:
• i) a pre-trained residual learning architecture comprising about 50 layers, wherein the residual learning architecture was pre-trained with at least one of images from an image database or a plurality of images of an ocular surface from a population of subjects;
• ii) batch normalization;
• iii) Dropout regularization;
• and
• iv) data augmentation prior to analysis by the residual learning architecture, wherein the data augmentation comprises data blending augmentation or data interpolating augmentation, which optionally is selected from at least one of mixup, random image flipping, random image rotation, or random image crops;
• and
• b) a backpropagation algorithm comprising:
• i) stochastic gradient descent, wherein the stochastic gradient descent further comprises momentum gradient acceleration with a value of 0.9;
• and
• ii) a learning rate of 0.00001.