Fig 1.
Automatic recognition of skin cancer is impeded by various factors.
(a) represents some interferences with important features (b) shows the similarity between benign and malignant lesions.
Table 1.
A summary of the literature review showing the past techniques and their limitations.
Fig 2.
Skin cancer ISIC archive dataset.
Table 2.
Dataset description.
Fig 3.
Three regions and artifacts in the skin lesion image.
Table 3.
Comparison of different approaches based on PSNR values.
Fig 4.
Steps performed for image pre-processing.
Fig 5.
Process flow of applying morphological closing.
Fig 6.
Resultant image of morphological closing with different kernel shape.
Fig 7.
Plots for three gamma correction states.
Fig 8.
Gamma corrected image transformation.
Fig 9.
Process flow of applying gamma correction.
Table 4.
Selected parameter values for morphological closing and gamma correction.
Fig 10.
Resultant image after applying Otsu’s thresholding.
Fig 11.
Changes of pixels after applying morphological dilation.
Fig 12.
Resultant mask after applying morphological dilation.
Fig 13.
Resultant mask after subtracting mask-1 from mask-2.
Fig 14.
Process flow of extracting the largest blob from mask-3.
Fig 15.
Process of extracting ROI by filling the largest blob.
Fig 16.
(a) Pre-processed images of skin lesion (b) Resultant images after ROI extraction.
Table 5.
Selected parameter value for segmentation process.
Fig 17.
Resultant image after applying the bilateral filter.
Fig 18.
Resultant image after applying Box blur Table 6 represents all the selected parameter values for the algorithms associated with down-scaling process.
Table 6.
Selected parameter for down-scaling process.
Table 7.
MSE, PSNR, SSIM, RMSE and DSC of 10 images.
Fig 19.
Histograms of original image vs box blur filter.
Fig 20.
Augmentation and classification process.
Table 8.
Number of images in classes after spitting.
Fig 21.
Architecture of proposed SCNN_12 model.
Table 9.
Performance analysis by changing batch size.
Table 10.
Performance analysis by changing output layer activation function.
Table 11.
Performance analysis by changing loss function.
Table 12.
Training loss(T_loss), training accuracy (T_acc), validation loss (V_loss), validation accuracy (V_acc), test loss (Te_loss), test accuracy (Te_acc) for each of the optimizers corresponding to the learning rates.
Table 13.
AUC, Recall (Rec), Specificity (Spe), Precision (Pre) and F1-score (F1) for each of the optimizers corresponding to the learning rates.
Fig 22.
Root mean squared error of all six configurations of SCNN_12.
Fig 23.
Mean absolute error of all six configurations of SCNN_12.
Table 14.
Configuration of optimal model based on hyper-parameters.
Fig 24.
Training and validation loss curve of 200 epochs for Adam with learning rate 0.001.
Fig 25.
Training and validation accuracy curve of 200 epochs for Adam with learning rate 0.001.
Fig 26.
Confusion matrix of the best configured model (Adam 0.001).
Fig 27.
ROC curve of best configured model (Adam 0.001).
Table 15.
Comparison among Bilateral and Bilateral + Box blur filter based on accuracy, average image size and average training time.
Fig 28.
Storage comparison for 10 images individually for both ‘DG-before’ and ‘DG-after’ where ‘DG-before’ indicates the dataset before applying down-scaling and ‘DG-after’ indicates the dataset generated after down-scaling.
Fig 29.
Training time per epoch comparison for 200 epochs of datasets ‘DG-before’ and ‘DG-after’ where ‘DG-before’ indicates the dataset before applying downscaling and ‘DG-after’ indicates the dataset generated after down-scaling.
Another impact, observed during training is that our down-scaled dataset (DG-after) records the highest accuracy within 200 epochs (Fig 30) whereas the non-down-scaled dataset (DG-before) takes more than 200 epochs to achieve the highest accuracy. This further adds to the efficiency of employing this method on future researches.
Fig 30.
Validation accuracy curve of both ‘DG-before’ and ‘DG-after’ datasets where ‘DG-before’ indicates the dataset before applying downscaling and ‘DG-after’ indicates the dataset generated after down-scaling.
Table 16.
Comparison of performance between geometric and photometric augmented dataset.
Table 17.
Result of 5-Fold cross-validation on augmented dataset.
Table 18.
Result of 10-Fold cross-validation on augmented dataset.
Table 19.
Result of 5-Fold cross-validation on the dataset before applying augmentation.
Table 20.
Result of 10-Fold cross-validation on the dataset before applying augmentation.
Table 21.
Performance of proposed model on the dataset after augmentation with various splitting ratios.
Table 22.
Performance of proposed model on the dataset before augmentation with various splitting ratios.
Table 23.
Performance of proposed model on noisy dataset.
Table 24.
Comparison of accuracy between the proposed system and existing systems.