Optimum machine learning models for osteosarcoma cancer detection and classification

Osteosarcoma is a bone-forming tumor which is more common with children and young adults than adults. Timely detection and classification of its type is crucial to its proper treatment and possible survival. Machine learning models, trained on datasets of the disease, are more effective detection and classification tool than hand-crafted features which are highly dependent on pathologists’ expertise. Publicly available raw osteosarcoma dataset was explored and preprocessed (including data denoising and data normalization). Three different datasets were then derived: the preprocessed dataset, and the preprocessed dataset with features selected via principal component analysis and a combination of analysis of variance and mutual information gain. Using the three datasets and eight machine learning (ML) algorithms, this study proposed three sets of optimum ML models (altogether 24 models) with their hyperparameters optimized using grid search. Then, the learned ML models were compared and validated using repeated stratified 10-fold cross-validation and 5 × 2 cross-validation paired t-test to select the best for our task. The ML model based on k-nearest neighbors algorithm proved to be the best, as it detected and classified osteosarcoma cancer in 344 ms with 100% Top-1 accuracy and F1- score and zero Type I and Type II errors. This performance exceeds those of existing algorithms for osteosarcoma cancer prediction. Thus, the proposed models are promising cutting-edge techniques for detecting osteosarcoma cancer to aid timely diagnosis, prognosis and treatment.