nnMobileNet: Rethinking CNN for Retinopathy Research

Abstract

Over the past few decades, convolutional neural networks (CNNs) have been at the forefront of the detection and tracking of various retinal diseases (RD). Despite their success, the emergence of vision transformers (ViT) in the 2020s has shifted the trajectory of RD model development. The leading-edge performance of ViT-based models in RD can be largely credited to their scalability-their ability to improve as more parameters are added. As a result, ViT-based models tend to outshine traditional CNNs in RD applications, albeit at the cost of increased data and computational demands. ViTs also differ from CNNs in their approach to processing images, working with patches rather than local regions, which can complicate the precise localization of small, variably presented lesions in RD. In our study, we revisited and updated the architecture of a CNN model, specifically MobileNet, to enhance its utility in RD diagnostics. We found that an optimized MobileNet, through selective modifications, can surpass ViT-based models in various RD benchmarks, including diabetic retinopathy grading, detection of multiple fundus diseases, and classification of diabetic macular edema. The code is available at https://github.com/Retinal-Research/NN-MOBILENET.

Figure 1.

Model size vs average performance (F1, Accuracy and AUC) on retinal multi-disease abnormal detection using RFMid dataset. Our method demonstrates superiority over other CNN/ViT based methods in terms of performance and efficiency.

Figure 2.

The roadmap of modifying a MobileNetV2 to the proposed no-new MobileNet (nnMobileNet) on the Messidor-2 dataset;

Figure 3.

The detailed architecture of the no-new MobileNet (Including the Channel configuration) and the inverted linear residual bottleneck used in the no-new MobileNet.

Figure 4.

Examples of data augmentation (Method III) and details of three sets of data augmentation we used.

Figure 5.

Empirical studies on Messidor-2 dataset where subpanel pictures (a), (b), (c), and (d) represent different experimental groups, each of which is independent of the others. D and SD-[x] in subpanel (b) denote Dropout and SpatialDropout in position [x] as shown in Fig.3(c), respectively.

Figure 6.

The comparative visualization on the Messidor-1 dataset was performed utilizing CAM [29]. We chose representative CNN/ViT-based methods with publicly available code, including MIL-VT [38], Swin-L [16], CrossFormer-L [35], CANet [15] and ReXNet [9].