3.3. Statistical Analysis

Our dataset contains skin lesion images from two categories, namely, benign skin moles and malignant skin moles. A total of 3297 images were sampled from the original ISIC 2018 challenge dataset from melanoma disease only. We used stratified probability sampling or randomization method to choose the images, where the samples were drawn for each stratum (group) at random but in proportional allocation. Furthermore, we adopted an unblinded technique, where randomization cannot be concealed.

As part of the statistical hypothesis testing, we used the corrected paired Student’s t-test to see whether the differences in the performances of various studied models are statistically significant or not. To accomplish this, the paired t-test was utilized using the null and alternative hypotheses stipulated earlier. Statistical significance was determined by p values less than 0.05. We made the following assumptions to use this method of testing regardless of whether there are significant differences between two sets of data. The measurements taken for one subject have no effect on the measurements taken for any other subject, and vice versa. Any of the paired measures must be acquired from the same subject. Finally, the observed performance differences follow a normal distribution. Both quantitative and qualitative results were provided to assess the effectiveness of the proposed models in detecting melanoma skin cancer in new patients. Among the quantitative data, accuracy and Cohen’s kappa are the most prominent. Qualitative results were supported by the suggested interpretability approach, which generates heatmaps to identify the parts of a melanoma image that are most suggestive of the disease. We used SciPy (1.8.0), which is a Python library for scientific computing, and is available for free and open source use.