Main Session

Sep 30

PQA 09 - Hematologic Malignancies, Health Services Research, Digital Health Innovation and Informatics

3631 - A Deep Radiomics Framework for Enhanced Prognostic Prediction in Locally Advanced Non-Small Cell Lung Cancer: A Multicenter Study

04:00pm - 05:00pm PT

Hall F

Screen: 5

POSTER

Presenter(s)

Runping Hou, PhD - Shanghai Chest Hospital Shanghai Jiao Tong University, Shanghai, Shanghai

R. Hou^1,2, W. Xia¹, M. T. Islam², X. Fu¹, and L. Xing²; ¹Department of Radiation Oncology, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China, ²Department of Radiation Oncology, Stanford University, Stanford, CA

Purpose/Objective(s):

Patients with unresectable locally advanced non-small cell lung cancer (LA-NSCLC) exhibits heterogeneous responses to standard concurrent chemoradiotherapy (CCRT), underscoring the need for precise risk stratification to optimize treatment strategies. We propose OmicsMap, a novel deep radiomics framework that leverages high-dimensional feature interactions to enhance prognostic prediction in LA-NSCLC. This multicenter study aimed to validate OmicsMap’s predictive capabilities using advanced deep radiomics techniques, facilitating personalized risk assessment and treatment decision-making.

Materials/Methods:

This retrospective study analyzed 389 LA-NSCLC patients treated with CCRT from two independent institutions. Patients were randomly assigned to a development cohort and a testing cohort with a ratio of 2:1. High-dimensional radiomic features were extracted from CT images. The selected discriminative features were then transformed into a structured two-dimensional representation, termed OmicsMap, in which data-specific feature interactions were embedded within a pixelated configuration. A convolutional neural network (CNN) was trained to extract deep features from OmicsMaps for progression-free survival (PFS) prediction. Model performance was assessed using time-dependent area under the receiver operating characteristic curve (AUC). Kaplan-Meier analysis was conducted to evaluate risk stratification efficacy.

Results: OmicsMap demonstrated superior prognostic performance, achieving time-dependent AUCs of 0.72, 0.75, and 0.71 at 1, 2 and 3 years in the independent testing cohort. Compared to the conventional Cox regression model built with tabular radiomics data (AUCs: 0.67, 0.65, 0.63), OmicsMap significantly improved predictive accuracy. The model effectively stratified patients into distinct risk subgroups with significantly different progression risks (P = 0.006, Hazard ratio = 0.573, 95% CI: 0.383–0.858) in the independent testing cohort.

Conclusion:

The proposed OmicsMap offers a novel approach to representing radiomic features and their interactions, enabling more accurate prognostic predictions for LA-NSCLC patients. This framework has the potential to assist clinicians in identifying high-risk individuals who may benefit from adjuvant therapeutic intensification.