Development of a Machine Learning algorithm for angle aperture grading from ophthalmic images

Glaucoma is an optic neuropathy, characterized by the gradual degradation of retinal cells, whose death is associated with an increased level of intraocular pressure. It is a primary cause of irreversible blindness, affecting more than 70 million people globally. Intraocular pressure is normally regulated by specialized tissues in the anterior chamber of the eye, along the interface between iris and cornea. The morphology of this region is clinically assessed by an ophthalmic examination, called gonioscopy, which allows to differentiate the two main sub-types of the pathology, i.e., angle-closure and open-angle glaucoma. In this work, a NIDEK GS-1 digital gonioscope is used to acquire a dataset of iridocorneal angle images showing different grades of an important morphological feature of the angle, i.e., its aperture. We investigated this classification task by developing and comparing Machine Learning algorithms consisting of three main phases: pre-processing, feature extraction, and classification. During the pre-processing step, the Bilateral filter is used to reduce noise, while the local contrast is improved by performing a Contrast Limited Adaptive Histogram Equalization. Subsequently, texture and shape features are extracted from all images. Two techniques, called Gray-Level Co-Occurrence Matrix and Gray-Level Run-Length Matrix, are used to compute texture descriptors. The iridocorneal angle shape descriptor is obtained through the Histogram of Oriented Gradients method. These features are employed to train four different models: Support Vector Machine, Random Forest, AdaBoost, and Gradient Boosting. Models' performance is, then, discussed and contextualized. Our results provide useful insights on potentials and limitations of Machine Learning systems in the context of automated gonioscopy, an important field of research currently almost unexplored.