A comparative assessment of machine learning models and algorithms for osteosarcoma cancer detection and classification

Abstract

Osteosarcoma is a bone-forming tumor that is more common in children and young adults than in adults. Timely detection and classification of its type is crucial to its proper treatment and possible survival. Machine learning (ML) models trained on disease datasets are more effective in detection and classification than the conventional methods with hand-crafted features highly dependent on pathologists’ expertise. A publicly available raw osteosarcoma dataset was explored and then preprocessed using different combinations of data denoising techniques (including principal component analysis, mutual information gain, analysis of variance and Kendall’s rank correlation analysis) and data augmentation to derive seven different datasets. Using the seven derived datasets and eight ML algorithms, this study designed and performed an extensive comparative analysis of seven sets of ML models (altogether over 160 models) with their hyperparameters optimized using grid search. The performance differences between the learned ML models were then validated using repeated stratified 10-fold cross-validation and 5x2 cross-validation paired t-tests to select the best model for our task. The empirical model based on the extra trees algorithm and fitted to class-balanced dataset via random oversampling and multicollinearity removed via principal component analysis proved to be the best, as it detected and classified osteosarcoma cancer in 10 ms with 97.8% area under the receiver operating characteristics curve and acceptably low false alarm and misdetection. Thus, the proposed models can be cutting-edge techniques for automated detection and classification of osteosarcoma tumors to aid timely diagnosis, prognosis, and treatment.