Abstract
Purpose :
Compliance in the usage of prescription eye drops is a major limitation in the management of ocular disease regardless of age, medical condition, or sightedness. However, eye drop medication adherence is particularly challenging in low-vision patients who may have difficulty in differentiating among medications because of small font, and similarities in size, color, and shape. Research has surveyed self-administration of medication in low-vision patients and found that a total of 89% of respondents were unable to read the prescription labels and 96% of these patients did not inform healthcare providers when they faced difficulties in handling their medication. The inability to correctly administer eye drops because of low-vision is a significant subset of eye drop medication non-compliance. We propose a deep learning algorithm which can accurately localize eye drop bottles to aid low-vision patients in recognition.
Methods :
Our image dataset was collected from the internet and spans the widely prescribed Latanoprost, Timolol, and Alphagan glaucoma eye drop bottle classes (1500 images). Bounding boxes are placed around the bottles in each image by a trained optometrist. Various image augmentations (sheer, zoom, contrast, flip) are applied. The YOLOv4 tiny computationally efficient object detection algorithm is trained on our annotated coordinate dataset.
Results :
The algorithm was trained across 300 epochs with a batch size of 16 and image size of 416x416. It recieved a final mean average precision (mAP) score at a 0.5 IoU threshold of 0.85 and at a 0.5:0.95 IoU threshold of 0.49. The algorithm recieved precision of 0.881 and recall of 0.924 across the test set (20% test split). Figure 1 shows box, objectness, and classification loss across epoch progression along with precision, recall, [email protected] and [email protected]:0.95 scores. Bottle localization inference on images averaged at 0.008 seconds on a P-100 GPU.
Conclusions :
We have proposed a custom deep learning object detection model which can accurately detect and localize three prescription eye drop bottles. With such a tool, patients can potentially identify and use bottles with more ease, lowering the amount of non-compliance. The proposed system is a proof-of-concept and more testing, a larger dataset, and classes are required before deployment onto a mobile phone.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.