Fig 1.
A typical colored fundus image.
Fig 2.
Research framework and workflow pipeline.
Fig 3.
Class label distribution for the study.
Table 1.
Image augmentation parameters.
Fig 4.
DIFE from the backbone architecture.
Fig 5.
PFMR module.
Fig 6.
Architecture of the LCC module.
Table 2.
Layer-wise specification of the LiteFeatNet model.
Table 3.
Comparison of number and size of parameters.
Table 4.
Target deployment platforms.
Fig 7.
(A) CNN-1 [17] (B) DenseNet121 (C) DeB5-XNet [24] (D) DIA-VXNET [29] (E) LightCNN [16] (F) MobileNet (G) MobileNetV2 (H) NAS_m1 [30] (I) NAS_m2 [30] (J) NASNetMobile (K) Proposed Model (LiteFeatNet) (L) Swin V2-Small (M) Swin V2-Tiny.
Fig 8.
(A) CNN-1 [17] (B) DenseNet121 (C) DeB5-XNet [24] (D) DIA-VXNET [29] (E) LightCNN [16] (F) MobileNet (G) MobileNetV2 (H) NAS_m1 [30] (I) NAS_m2 [30] (J) NASNetMobile (K) Proposed Model (LiteFeatNet) (L) Swin V2-Small (M) Swin V2-Tiny.
Fig 9.
Confusion matrix for LiteFeatNet using RFMiD.
Table 5.
Overall performance metrics for RFMiD.
Table 6.
Class-wise performance metrics for RFMiD.
Fig 10.
Class label distribution for RFMiD 2.0.
Fig 11.
Pre-processing pipeline for RFMiD 2.0.
Fig 12.
Evaluation mechanism for RFMiD 2.0.
Fig 13.
Confusion matrix for LiteFeatNet using RFMiD 2.0.
Table 7.
Evaluation metrics for RFMiD 2.0.
Fig 14.
External Dataset for Scalability Assessment.
Table 8.
Class labels and the number of images supplemented for scalability assessment.
Table 9.
Performance metrics for 4 retinal pathologies.
Table 10.
Performance metrics for 5 retinal pathologies.
Table 11.
Ablation study experiments for the proposed LiteFeatNet.
Fig 15.
Performance evaluation of ablation experiments.
Table 12.
Ablation experiments for other backbone architectures.
Fig 16.
Performance improvement with NASNetMobile backbone.
Fig 17.
Performance improvement with DenseNet121 backbone.
Fig 18.
Performance improvement with MobileNetV2 backbone.
Fig 19.
Comparison of inference times.