Multilayer Perceptron Analysis of Radiomics to Predict Local Recurrence of Lung Cancer After Radiotherapy

Abstract

Purpose: to assess the likelihood of local recurrence of lung malignancies following stereotactic ablative radiotherapy (SABR) by evaluating clinical and radiomic features with machine learning and novel use of deep learning methods.

Methods: pre-treatment CT images were attained from seventy patients with primary lung malignancies. The malignancy was segmented by the treating radiation oncologist and 107 radiomic features were extracted from the image. The data underwent feature reduction via Spearman’s correlation and selection with adapted LASSO regression analysis. A random forest model and a multilayer perceptron (MLP) with cost-sensitive classifier were independently used to assess for local recurrence of malignancy. The recurrence likelihood predictions from each of these were used to stratify patients into groups with high and low risk of recurrence. These were assessed for time-to-event predictions using Kaplan-Meier analyses and Gray’s test to evaluate the separation between the high and low-risk groups. The prognostic capacity of the models was evaluated with a concordance index, 95% confidence intervals and bootstrapping (10,000 iterations).

Results: the MLP was able to predict the recurrence of malignancy with 100% sensitivity and 91% specificity (AUC 0.95). The MLP predictions showed statistically significant separation of high and low-risk patients, and robust model fit (p=0.04, c=0.79), which out-performed random forest model predictions (p=0.15, c=0.41) that did not reach statistical significance.

Conclusions: radiomic data analysis with an MLP showed improved prediction potential within this dataset compared to random forest models for predicting local recurrence of lung cancer.