Room: Davidson Ballroom B
Purpose: Identifying potentially malignant lung nodules in CT is essential for screening and diagnosis of lung cancer. Radiomics has shown promising result in nodule classification. Since not all the radiomic features play a positive role in building the predictive model, and some of them may be redundant or even reduce the model performance, selecting an optimal feature subset is critical. We aim to develop a new multi-objective based feature selection (MO-FS) algorithm for radiomics-based lung nodule malignancy classification.
Methods: The proposed MO-FS method is improved from an iterative multi-objective immune algorithm (IMIA), in which both sensitivity and specificity are considered as the objective functions simultaneously during the feature selection. The improvements of MO-FS include: 1) a modified entropy based termination criterion (METC) so that the algorithm stops automatically rather than relying on a preset fixed number of generations; and 2) a selection methodology for multi-objective learning using the evidential reasoning approach (SMOLER) to select the optimal solution from Pareto-optimal set automatically. The MO-FS consists of two phases: 1) generating Pareto-optimal set; and (2) SMOLER based optimal solution selection. The first phase includes initialization, clonal operation, adaptive mutation operation, updating solution set and IETC based termination. The dataset used in this study includes 431 malignant nodules and 795 benign nodules extracted from LIDC-IDRI. Two hundred fifty-seven features including intensity, texture and geometry were extracted from CT images.
Results: Ten times running results of 5-folder cross-validation (meanÂ±standard deviation) of area under the curve (AUC), accuracy, sensitivity, and specificity for MO-FS are 0.9349Â±0.0017, 0.8899Â±0.0014, 0.9068Â±0.0043, 0.8587Â±0.0057, respectively. It outperforms the sequential forward selection (SFS) method that achieves AUC of 0.9055.
Conclusion: We developed a new MO-FS algorithm for radiomics-based lung nodule malignancy classification. The model built on MO-FS selected features obtained better performance compared with other methods such as SFS.
Not Applicable / None Entered.
IM/TH- Image Analysis (Single modality or Multi-modality): Machine learning