Machine learning techniques for ultrasound biomicroscopy images for accurate automatic determination of phakic status

Abstract

Purpose : Ultrasound biomicroscopy (UBM) is a point of care, cost-effective, and non-invasive method for assessing anatomical features of the anterior segment. Previous studies have demonstrated the potential of automated structure classification in UBM images, but these studies have not yet explored lens or lens-related structure classification. Lens status classification is an objective, ground truth task that necessitates identification of the lens and adjacent anatomy. We developed and validated a proof of concept convolutional neural network with transfer learning to classify lens status from UBM images.

Methods : We collected 285 UBM images from 66 pediatric and adult patients between ages 3 weeks – 88 years old with known lens status classified as phakic, pseudophakic, or aphakic. We split patients and their associated images into testing, validation, and training data. A pretrained model, Densenet-121, with a fully connected custom classifier was trained on these images. Weighted-average precision, recall, and F1-score were calculated for the classifier’s output on the testing dataset with 5-fold cross validation. We applied gradient-weighted class activation mapping to generate a heat map and visualize object focus in the network’s final layer.

Results : Our neural network trained across 60 epochs achieved a recall of 95.79%, a precision of 95.71%, and an F1-score of 95.71%. Feature saliency heat maps of the network’s final layer consistently involved the lens or adjacent structures.

Conclusions : Preliminary results demonstrate that a neural network with transfer learning trained on UBM images can classify lens status with satisfactory recall and precision and encouraging feature saliency suggestive of lens localization. Further studies will be needed to explore the role of automated classification in UBM images of more complex anterior segment pathology.

This is a 2020 Imaging in the Eye Conference abstract.

 

Input images (left) and associated feature saliency heat maps (right) for correctly identified (a) phakic, (b) pseudophakic, and (c) aphakic images.

 

Confusion matrix of the model's classification on all testing set data aggregated across 5-fold cross validation. Element (x,y) of this matrix represents the number of UBM images predicted as "x" lens status with a true lens status, "y".