Abstract
Purpose :
Sjögren syndrome (SS) can lead to severe ocular pain and discomfort. Patients with SS exhibit corneal inflammation and alterations of the sub-basal plexus nerves. Additionally, SS is hard to diagnose as it requires invasive exams (e.g. salivary glands biopsy). In Vivo Confocal Microscopy (IVCM) offers to assess corneal nerve fibers in non-invasive way. Analyzing such images with Deep Learning based method could help to identify patterns and redundancies so far invisible to ophthalmologists.
Our solution offers to use Convolutional Neural Networks (CNN) in a Multiple Instance Learning setting to help diagnose Sjögren syndrome from IVCM images.
Methods :
Material:
Cohort of 80 patients from 15-20 hospital (17 healthy volunteers, 63 patients with SS), splitted between the training set and the test set.
The dataset consists of a hundreds of IVCM images per patient, acquired and pre-selected by the ophthalmologist and clinical data. See figure 1.
Method:
The method relies on two steps:
1) Segmentation of nerves, inflammatory cells and neuromas was performed on 200 IVCM images using a U-Net network. Abstract features learned from the segmentation task were then transferred to a second network (transfer learning).
2) A second network was used to perform the diagnostic task. It takes advantage of the Multiple Instance Learning (MIL) framework to let the model benefit from the multiple images available for each patient. This network relies on two parts: the “tile predictor” attributing a score to each image, and the “aggregator” (attention mechanism was used) to aggregate the score of each images in the best way possible and perform a good diagnosis.
Results :
Figure 2 presents the confusion matrix for the binary classification on control patients versus SS patients. The model demonstrated an accuracy of 81.1 % and an average ROC AUC of 0.69, as well as the ROC Curve for the SS class on the test set.
Conclusions :
This model demonstrates promising results in the ability to diagnose and monitor Sjögren syndrome from IVCM images, using non-invasive exams. Future works will consist in validating and improving the current algorithm on a larger cohort of patients.
This is a 2020 ARVO Annual Meeting abstract.