Table 1.
Summary table of previous state-of-the-art works.
Fig 1.
Workflow of the proposed system.
First, data collection, then data preprocessing, including augmentation and Gaussian filter, followed that applying deep learning models, and finally, result analysis.
Fig 2.
Samples of arsenic infected and normal skin images.
Table 2.
Data augmentation and resize values for this study.
Fig 3.
Architecture of the proposed model. Xception as the backbone model and added the Inception module with it.
Fig 4.
Block diagram of the inception module with their kernel size and number of filters for each layer.
Table 3.
Hyperparameters for proposed model.
Table 4.
Elaboration of the setup for the research.
Fig 5.
Confusion matrices for InceptionV3 and VGG19.
Here, 0 = infected, 1 = normal.
Fig 6.
Confusion matrices for EfficientNetV2B0 and ResNet152V2.
Here, 0 = infected, 1 = normal.
Fig 7.
Confusion matrices for Xception and Proposed model.
Here, 0 = infected, 1 = normal.
Fig 8.
ROC curves for InceptionV3 and VGG19.
Here, 0 = infected, 1 = normal.
Fig 9.
ROC curves for EfficientNetV2B0 and ResNet152V2.
Here, 0 = infected, 1 = normal.
Fig 10.
ROC curves for Xception and Proposed model.
Here, 0 = infected, 1 = normal.
Fig 11.
Proposed model’s training performance on 80:10:10 data split: accuracy vs. epochs and loss vs. epochs.
Table 5.
Performance metrics comparison of various models for 80:10:10 data split. Here, 80% for training set, 10% validation set, 10% test set, test accuracy as Test_Acc.
Fig 12.
Training accuracy and loss of the various deep learning models based on 80:10:10 data split.
Fig 13.
Grad-CAM visualization (a) and (c) input image, (b) and (d) Grad-CAM overlay on original image. Here, 0=infected, 1=normal.