Introduction Malignant occurrences are increasing in frequency, and skin cancer is emerging as a significant public health concern. Conventional approaches, encompassing comprehensive procedures such as consulting expert opinions, tend to decelerate the therapeutic process. Methods A method for classifying skin cancer that incorporates both manually extracted and automatically derived features. This study introduces (i) an innovative method for extracting the feature known as Border Irregularity, which significantly enhances detection capabilities. (ii) This investigation demonstrates that Fitzpatrick skin type significantly influences the detection of cancer. (iii) The Dual Stream Residual Squeeze Excite Network (DSRSENet), a novel deep learning architecture, is employed to extract non-handcrafted features. (iv) Handcrafted features are filtered prior to concatenation with non-handcrafted features through Self Attention-Mutual Information (SAMI) method, which operate based on weighted ranking. (v) A unique optimal feature selection (OFS) method is employed to enhance detection accuracy on concatenated features. Results This empirical study is conducted on two publicly available datasets, PADUFES₂0 and HAM10000, to ensure generalization and robustness. Only handcraft features exhibit detection accuracy of 67. 39% and 84. 27%, improved to 73. 80% and 89. 42% with the application of SAMI on both datasets, respectively. After concatenation with the non-handcraft features, OFS further improved the detection accuracy to 93. 26% and 95. 66% for both datasets respectively, by tuning the hyperparameters precisely. Discussion The results demonstrate that Fitzpatrick Skin Type and the proposed Border Irregularity extraction method have a significant impact on skin malignancy detection. Furthermore, the proposed DSRSENet model effectively captures non-handcrafted features, while the SAMI-based filtering and OFS method contribute to improved feature selection. The integration of handcrafted and non-handcrafted features results in superior performance compared to existing methods.
Reshma et al. (Wed,) studied this question.