A review and comparative study of cancer detection using machine learning: SBERT and SimCSE application

Background: Pretrained convolutional neural networks and traditional machine learning algorithms have extensively been considered for the detection of various cancers using image data and clinical data as the input source. Initially, a series of preprocessing steps and image segmentation steps are performed to extract region of interest (ROI) features from noisy features. Then, the extracted features are applied to several machine learning and deep learning methods for the detection of cancer. Method: In this work, a review of all the methods that have been applied to develop machine learning algorithms that detect cancer is provided. This study solely focuses on work that investigates four cancer types (lung, breast, prostate, and colorectal cancer) that have been published in the last five years (2018-2022). Then, a colorectal cancer detection framework is proposed using colorectal cancer DNA sequences as the only input source. For data representation, recent state-of-the-art sentence transformers namely: Sentence-BERT (2019) and the unsupervised SimCSE (2021) are applied. The novelty of this approach is that, the two sentence transformers have not been used to represent DNA sequences of matched tumor/normal pairs for cancer detection using machine learning. For classification, the learned DNA representations were then provided as input to traditional machine learning algorithms. Result: Using SBERT sentence embeddings, the XGBoost model achieved the best overall, achieving the highest accuracy of 73 ± 0.13 % on the development set, and 73.3 ± 0.20 % on an independent test set. With regard to the SimCSE embeddings, the Random forest model achieved the best overall on the development set with an accuracy of 71.6 ± 1.47 %. And on the test set, the XGBoost model achieved the highest accuracy of 75 ± 0.12 %. In summary of these findings, no significant improvement was observed in the performance of the machine learning models when using sentence representations from either SBERT or SimCSE transformer model. cancer detection, machine learning, SentenceBert, SimCSE