WEDNESDAY, Dec. 20, 2023 -- A proposed blending ensemble-based deep learning model (BEDLM-CMS) can be used for early diagnosis of cutaneous melanoma, according to a study published online Dec. 14 in Scientific Reports.

Noting that the accuracy of the state-of-the-art computer-aided diagnosis systems is unsatisfactory and that the major drawback of medical imaging is the shortage of labeled data, Asghar Ali Shah, from Bahria University in Islamabad, Pakistan, and colleagues proposed a BEDLM-CMS to detect mutation of cutaneous melanoma by integrating long short-term memory (LSTM), bidirectional LSTM (BLSTM), and gated recurrent unit (GRU) architectures. The dataset included 2,608 human samples and 6,778 mutations with 75 types of genes. The genes that function as biomarkers for early diagnosis and prognosis most prominently were utilized; the most prominent features were extracted. Different DL models optimized through grid search techniques were applied to diagnose melanoma. Result validity was confirmed using several techniques.

The researchers found that in the independent set test, BEDLM gave the highest accuracy of 97 percent compared with 94 and 93 percent, respectively, in the self-consistency test and tenfold cross-validation test. The accuracy rates of the self-consistency test, independent set test, and tenfold cross-validation test were 96, 94, and 92 percent, respectively, for LSTM; 93, 94, and 91 percent, respectively, for GRU; and 99, 98, and 93 percent, respectively, for BLSTM.

"In the future this technique can be used in identifying other life-threatening diseases and other deep learning models can be used for obtaining more accuracy and efficiency," the authors write.