Frisén Grade Categorization in Papilledema using Vision Transformers

Abstract

Purpose : In previous work, we applied DenseNet architecture for classifying papilledema severity. This model correctly estimated swelling severity within one Frisén grade in 71.8% of cases. The present study explores the potential of Vision Transformers (ViT) as a novel methodology to discern papilledema from normal ocular fundus images and to classify the stages of papilledema severity more accurately. We hypothesize that ViT could improve the accuracy of automated Frisén grade classification for papilledema.

Methods : We adapted a pre-trained ViT model, RETFound, initially trained on a corpus of 1.6 million retinal images, for papilledema classification. Our dataset encompassed 5,908 fundus photographs from 165 patients in the Idiopathic Intracranial Hypertension Treatment Trial. After removing backgrounds and resizing images to 256x256 pixels, the dataset was split into 80% for training and the remainder for testing, with a portion of the training set (20%) earmarked for validation. The RETFound model was restructured to include a new head, enabling binary and six-class classification tasks.

Results : In binary classification, the ViT model exhibited an average accuracy of 99.76% and an F1-score of 99.87%. When applied on multi-class classification for Frisén grades, the model attained an average accuracy of 93.1% and an F1-score of 81.40%. Notably, the classification of grade 5 was less accurate when compared to the other grades, with only 66.67% of true grade 5 cases being correctly identified; the remaining were misclassified as grade 4, a variance ascribed to the limited representation of grade 5 within the dataset.

Conclusions : The RetFound ViT model has substantiated its capability to accurately discern and classify papilledema in fundus photography, achieving significant improvement when compared to the DenseNet. Furthermore, the ViT model has consistently delivered considerable accuracy throughout various papilledema grades. Future directions include applying this model to regression tasks for continuous severity grading, incorporating activation maps to identify defining features for each Frisén grade, and incorporating this output into larger fusion models, thus enhancing the model's diagnostic interpretability and clinical utility.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 2 - Confusion Matrix for Frisén Grade Classification

 

Figure 1 - Confusion Matrix for Binary Classification