Evolutionary ensemble feature selection learning for image-based assessment of lymphedema arm volume

Arm lymphedema is a side effect of surgery and radiation therapy in patients with breast cancer. There are several methods to measure lymphedema arm volume, such as water displacement (WD) and circumferential measurement (CM). Although these techniques have a satisfactory level of accuracy, they have some limitations for clinical use in practical applications. To overcome these drawbacks, a horizontal-vertical image scanning (HVIS) tool, is presented as a valid method for arm volume measurement. To predict the arm volume, an evolutionary ensemble feature selection learning (named EEFSL) comprising different base learners (KNN, SVM, ANFIS, Decision Tree, and Naive Bayes) is presented. In order to improve the performance of the EEFSL model, its hyperparameters are optimized using genetic algorithm (GA). The hyperparameters include the weights of different base learners and the specific features for each base learner. The fitness function of GA is to maximize the correlation between predicted and actual arm volumes. The proposed method was successfully developed for arm measurement of 60 arms from 30 women with arm lymphedema. Obtained results indicate that the proposed HVIS-EEFSL method is a reliable and valid technique to measure arm volume of patients with lymphedema, suitable for daily clinical use instead of WD and CM.