Fig 1.
Visual examples depicting the seven categories of pigmented skin lesions.
Fig 2.
MultiRes block: The rounded rectangle represents a concatenation operation where the black block represents a 3 × 3 convolution, the green block represents a 5 × 5 convolution and the red one represents a 7 × 7 convolution.
Finally a skip connection is added along with 1×1 filter.
Fig 3.
Res path: The encoder features are passed through a series of convolutions instead of linearly connecting them to the decoder features.
Fig 4.
Bayesian MultiResUNet comprises an encoder and a decoder pathway, with skip connections and dropout layers between the corresponding layers in pooling and upsampling blocks.
Fig 5.
We clearly observe the Bayesian MultiResUNet outperform the U-Net with far more precise boundaries.
Table 1.
Comparitive study of various segmentation models on the ISIC 2018 Task-1 dataset.
Fig 6.
Uncertainty estimation of the Bayesian MultiResUNet.
Table 2.
Comparitive study of classification models on the ISIC 2018 dataset.
Table 3.
Comparitive study of Bayesian versions of top two classification models on the ISIC 2018 dataset.
Table 4.
Class wise performance of Bayesian DenseNet-169.
Fig 7.
Posterior probability distributions for each of the possible scenarios i.e incorrect-uncertain (iu), correct-uncertain (cu), correct-certain (cc), and incorrect-certain (ic).
Assuming that the combined φTis 0.35, the red region indicates the posterior probability distribution for the incorrect class where as the green region indicates the posterior probability distribution of the correct class.
Table 5.
Comparison of different uncertainty types.
Fig 8.
Top regions of interest identified by XRAI for classification made using the Bayesian DenseNet-169 which is part of our SkiNet framework.
Fig 9.
Model’s region of interest depicted by XRAI and Guided Grad CAM.
For the skin lesion in Fig 9a we observe that the bottom right part of the skin lesion depicted in the XRAI heatmap in Fig 9b and Guided Grad CAM map in Fig 9c is of importance to our model. This is clearly depicted by the top 10% and top 5% plots in Fig 8c and 8d, showing that the model was heavily influenced by the dark red region. From Fig 9e, we observe that the top right part of the lesion is of importance to the model. Fig 8g and 8h show that the model is influenced by the reddish pinkish region present in the top right part of the skin lesion.
Table 6.
Comparative analysis between different explainability techniques.
Table 7.
Categorical segregation of predictions made on our test data.
Fig 10.
Image received by the Classification Algorithm (a) Image passed into stand-alone DesNet-169 (b) Image passed after segmentation step of SkiNet framework.
Fig 11.
Regions of Interest identified by XRAI a.) Before Segmentation b.) After Segmentation.
Table 8.
Comparitive analysis between the performance of a Stand-alone DenseNet-169 and the SkiNet pipeline of image in above figure [CU →CC].
Fig 12.
Image received by the Classification Algorithm (a) Image passed into stand-alone DesNet-169 (b) Image passed after segmentation step of SkiNet framework.
Fig 13.
Regions of Interest identified by XRAI a.) Before Segmentation b.) After Segmentation.
Table 9.
Comparitive analysis between the performance of a Stand-alone DenseNet-169 and the SkiNet pipeline [IU →CC] of above image.
Fig 14.
Image received by the Classification Algorithm (a) Image passed into stand-alone DesNet-169 (b) Image passed after segmentation step of SkiNet framework.
Fig 15.
Regions of Interest identified by XRAI a.) Before Segmentation b.) After Segmentation.
Table 10.
Comparitive analysis between the performance of a stand-alone DenseNet-169 and the SkiNet pipeline of above image [IC →CC].