Predicting Survival and Recurrence of Lung Ablation Patients Using Deep Learning-Based Automatic Segmentation and Radiomics Analysis.

Abstract

Purpose: To predict survival and tumor recurrence following image-guided thermal ablation (IGTA) of lung tumors segmented using a deep learning approach.

Methods and materials: A total of 113 patients who underwent IGTA for primary and metastatic lung tumors at a single institution between January 1, 2004 and July 14, 2022 were retrospectively identified. A pretrained U-Net model was applied to the dataset of pre- and post-procedure CT scans to segment lung zones. Following lung segmentation, a U-shaped encoder-decoder transformer architecture (UNETR) was trained to segment lung tumors from pre- and post-procedure CT scans, and radiomic features were automatically extracted. These features were input into a support vector machine (SVM)-based survival prediction model trained to assign rank scores to samples based on binary survival or recurrence label and follow-up time. C-index and time-dependent AUC were subsequently calculated to evaluate model performance.

Results: Initial tumor segmentation using UNETR achieved a Dice score of 0.75. Applying a radiomics-based survivability prediction model to the post-procedure scans resulted in a c-index of 0.71 and a time-dependent AUC of 0.75. In contrast, when this model was applied to pre-procedure scans, it achieved a 0.56 for both metrics. For predicting time to recurrence, the radiomics-based model achieved a c-index of 0.65 and a time-dependent AUC of 0.72 on post-procedure imaging. In contrast, when this model was applied to pre-procedure scans, it achieved a 0.54 for both metrics.

Conclusion: Radiomic feature analysis of lung tumors following automatic segmentation by a state-of-the-art transformer-based U-NET may predict survival and recurrence following image-guided thermal ablation of pulmonary malignancies.

Level of evidence: Level 3, Retrospective cohort study.