A Low-Cost High-Performance Data Augmentation for Deep Learning-Based Skin Lesion Classification

Abstract

Objective and Impact Statement. There is a need to develop high-performance and low-cost data augmentation strategies for intelligent skin cancer screening devices that can be deployed in rural or underdeveloped communities. The proposed strategy can not only improve the classification performance of skin lesions but also highlight the potential regions of interest for clinicians' attention. This strategy can also be implemented in a broad range of clinical disciplines for early screening and automatic diagnosis of many other diseases in low resource settings. Methods. We propose a high-performance data augmentation strategy of search space 10¹, which can be combined with any model through a plug-and-play mode and search for the best argumentation method for a medical database with low resource cost. Results. With EfficientNets as a baseline, the best BACC of HAM10000 is 0.853, outperforming the other published models of "single-model and no-external-database" for ISIC 2018 Lesion Diagnosis Challenge (Task 3). The best average AUC performance on ISIC 2017 achieves 0.909 (±0.015), exceeding most of the ensembling models and those using external datasets. Performance on Derm7pt archives the best BACC of 0.735 (±0.018) ahead of all other related studies. Moreover, the model-based heatmaps generated by Grad-CAM++ verify the accurate selection of lesion features in model judgment, further proving the scientific rationality of model-based diagnosis. Conclusion. The proposed data augmentation strategy greatly reduces the computational cost for clinically intelligent diagnosis of skin lesions. It may also facilitate further research in low-cost, portable, and AI-based mobile devices for skin cancer screening and therapeutic guidance.

Figure 1

BACC performance of EfficientNets trained adopting the best LCA strategy ($n=5$), ImageNet-based AutoAugment strategy ($n=5$), and General Augmentation strategy ($n=5$). The DCNNs from left to right on the $x$-axis correspond to EfficientNet b0-b7.

Figure 2

Performance of (a) BACC, (b) average AUC ($n=7$), (c) average specificity ($n=7$), (d) average accuracy ($n=7$), and (e) average precision ($n=7$) of EfficientNets trained adopting the searched augmentation strategy. The DCNNs from left to right on the $x$-axis correspond to EfficientNet b0-b7.

Figure 3

Loss curve of (a) train data and (b) test data on ISIC 2017 dataset. Here, the DCNNs from left to right on the $x$-axis correspond to EfficientNet b0-b7.

Figure 4

Heatmaps of lesions in different types in (a) HAM10000, (b) ISIC 2017, and (c) Derm7pt dataset.

Figure 5

Work flow for the two-stage approach to search for the best combination of network and augmentation strategy.

Figure 6

Representations of augmentation effects of different substrategies in the probability of 0.5.