Abstract
Purpose :
To compare Gradient class activation maps (Grad-CAM) heatmaps of three Convolutional Neural Networks (CNNs) in localizing glaucomatous optic neuropathy (GON) and to evaluate their clinical usefulness
Methods :
The data used for this project consisted of 193 color retinal images including the Optic Nerve Head (ONH) collected with the Compass fundus perimeter (Centervue, Italy), for a total of 101 images labelled as healthy (NRM) and 92 as glaucoma (GLC). The data were processed to obtain different crop sizes as input for three CNNs: Square Crop (SC), Halfmoon Crop (HC) and Fovea-Extended Crop (FEC) models, respectively using a 200x200, 700x300 and 700x500 pixel input crop. Grad-CAM Heatmaps (GCH) were generated for the three CNNs (example shown in Fig1a). Last convolutional layer (CL) and average GCH were analyzed for all CNNs to understand the decision-making process. A qualitative approach was adopted to classify GCH into 4 pre-defined categories. Fig1b provides a visual explanation for each category. The percentage of activation (PA) of each category was compared among CNNs
Results :
The results are reported in Fig2. For images predicted as GLC by the CNNs, GCH should be active on image portions involving ONH, which are clinically meaningful for GON. This is supported by the high PA of GCH in the category involving ONH for the HC model. On the other hand, for images predicted as NRM, GCH should not highlight any image area. However, the results show a low percentage of GCH in the category “no activation” and a high percentage in the category involving crop edges and ONH
Conclusions :
The results indicate that GCH are reasonable for images predicted as GLC, whereas they may not be clinically useful for images predicted as NRM. HC model shows the higher PA on image portions involving ONH, suggesting that this is the most clinically meaningful model. In conclusion, GCH of HC model could be useful in a glaucoma-care clinical setting to better understand the CNN outcome
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.