Author: Ryan Alden, Tithi Biswas, Kaushik Halder, Felix Maria-Joseph, Michael Mix, Rihan Podder, Tarun Kanti Podder 👨🔬
Affiliation: SUNY Upstate Medical University, IIT-Roorkee, University of Florida 🌍
Purpose: Early-stage NSCLC patients undergoing SBRT often die due to intercurrent illnesses. However, prediction of overall survival (OS) remains crucial due to the risk of disease recurrence. This study aims to design and validate an OS prediction model based on convolution neural network (CNN) while incorporating the radiomics features gleaned from planning CT.
Methods: Planning CTs of 171 patients who underwent SBRT were used for this study. CT image preprocessing involved three steps: masking based on tumor volume, reshaping the masked image to standardized dimensions, and normalization to a uniform scale. Relevant radiomics features were extracted from the PTVs contoured by radiation oncologists for treating the patients. The CNN model was designed based on three residual blocks, each consisting of a convolution layer, ReLU activation layer, Batch Normalization layer, and Max pooling layer. Before feeding the data to a dense neural network, two flatten layers were used. The dense neural network consists of four hidden layers. Best ten slices were selected for the inputs to the CNN model for each patient based on the standard deviation of pixel intensity values. Performance of the developed model was compared to two commonly used pre-trained prediction models, i.e., VGG16 and ResNet50.
Results: Efficacy of the designed CNN-based 2-year OS-prediction model was evaluated using 10 times bootstrap splitting with 80% data for training and 20% for testing. CNN-based model achieved ROC-AUC 0.74 (±0.04), sensitivity 0.73 (±0.15), and specificity 0.74 (±0.16). Whereas VGG16 and ResNet50 models yielded ROC-AUCs 0.68 (±0.03) and 0.65 (±0.03), respectively. The CNN-based prediction model outperformed the existing pretrained models by 8.8% (VGG16) and 13.8% (ResNet50) for ROC-AUC metric.
Conclusion: This study demonstrates enhanced performance of our developed CNN-based prediction model for NSCLC patients’ OS after SBRT. Further study with a larger cohort of patients from multiple institutes is warranted.