Deep neural network based prognostic modelling for lung cancer utilising radiomics and clinical data

The purpose of this research paper was to assess the use of deep neural networks in prognostic stratification of lung cancer by combining clinical data and Radiomics features. The study included a dataset of 422 non-small cell lung cancer (NSCLC) patients with tumor segmentation and survival status.

Radiomics feature extraction was performed on the CT datasets of the patients, and a two-step cluster analysis was conducted on the radiomic features to identify predictors with significant importance for survival probability. These predictors, along with clinical features, were used as input nodes for an Artificial Neural Network (ANN) model. The dataset was divided into training and validation cohorts, and the model's accuracy was evaluated using ROC analysis.

The survival analyses based on clinical features alone did not yield significant results. However, the two-step cluster analysis based on radiomic features successfully segregated the dataset into two clusters with significantly different median overall survival (OS). The ANN model trained on 70% of the dataset achieved an accuracy of 73.2% in predicting survival probability. When applied to the validation cohort, the model's accuracy increased to 80%. The final ANN model was validated using ROC analysis, which showed an Area Under Curve (AUC) of 0.84.

The study concluded that combining radiomic and clinical data with an ANN model can provide a proof-of-concept for predicting patient outcomes in lung cancer. This demonstrates the potential application of artificial intelligence in prognostic modeling using a combination of radiomic and clinical information.