Deep transfer learning for automated liver cancer gene recognition using spectrogram images of digitized DNA sequences

Abstract

DNA biomarkers are considered to be an important diagnostic factor for early detection of cancer. With the development of technology, many machine learning methods have been introduced to detect cancer biomarkers from DNA sequences in the literature. In this study, a new approach was proposed for the detection of cancer genes, which is an important step for the prediction of cancer. In the proposed approach, the gene sequences were digitized by three different numerical mapping techniques. Following digitization, these DNA sequences were initially examined with two different ways as one-dimensional signal and two-dimensional spectrogram images. Firstly, the digitized sequences were examined with the designed CNN model as a one-dimensional signal. Secondly, DNA signals were converted to 2D spectrogram images and examined with two different 2D CNN models. In the first model, feature vectors were obtained by VGG16 and classified by SVM. In the second model, new layers were added to the final output layers of VGG16, and fine-tuning was applied. An accuracy of 80.36% was obtained in one-dimensional CNN model, an accuracy of 98.86% was achieved in the model where features were extracted with VGG16 and classified with SVM, and an accuracy of 100% was observed in the model where fine tuning was applied to the layers of VGG16. The proposed method indicated that effective features were extracted with CNN models to distinguish liver cancer and normal liver gene sequences. The application results showed that the system is ready to be tested with a larger dataset and different cancer types.