Early identification of epidermal growth factor receptor (EGFR) mutations is critical for selecting therapy in patients with non-small cell lung cancer (NSCLC). Concurrent tumor protein 53 (TP53) mutations are linked to unfavorable clinical outcomes in advanced NSCLC patients treated with EGFR-tyrosine kinase inhibitors (TKIs). Conventional molecular testing through surgical resection or needle biopsy remains the gold standard but is limited by invasiveness, sampling bias, and cost. To address these challenges, we propose a pre-trained three-dimensional convolutional neural network (3D CNN)- based framework for non-invasive prediction of solitary EGFR mutations and concurrent EGFR/TP53 co-mutations directly from chest CT scans. We employ a pure 3D CNN based on the ResNet3D-18 architecture, using only 3 × 3 × 3 convolutional kernels, and initialized with pre-trained weights from the large-scale Kinetics-400 dataset. The model is fine-tuned rather than trained from scratch, enabling transfer of generic spatial–temporal representations to the medical imaging domain and mitigating overfitting on a limited dataset. The network takes a two-channel volumetric input composed of the CT volume and a manually segmented tumor mask, which are concatenated along the channel dimension to provide complementary anatomical and lesion-focused information. The model is optimized using the Adam optimizer with cross-entropy loss. Model performance is evaluated using five independent random train–test splits to assess robustness and stability. The retrospective cohort consisted of 89 patients, with five stratified random splits into training and test sets. Across splits, the pre-trained 3D CNN demonstrated robust and reproducible discriminative performance for identifying EGFR and EGFR/TP53 mutation status, achieving a mean area under the curve (AUC) of 0.93 ± 0.04 95% confidence interval (CI) was 0.88–0.98 on test sets. Compared with a Support Vector Machine (SVM) and an identical ResNet3D-18 trained from random initialization, the pre-trained model consistently showed improved performance across accuracy, recall, and AUC metrics. Performance was comparable to medical-domain pre-training approaches (e.g., Med3D), indicating that effective volumetric representations can be learned through different pre-training strategies. Our results suggest that pre-trained 3D CNNs provide a potentially stable and data-efficient approach for radiogenomic prediction in lung adenocarcinoma under limited sample conditions. Rather than replacing molecular testing, the proposed framework may serve as a non-invasive adjunctive tool to support mutation status assessment. However, given the relatively small, single-center cohort, these findings should be interpreted as preliminary and hypothesis-generating, and the observed performance may partially reflect characteristics of the study dataset. Larger multi-center studies with external validation are required before clinical translation.
Cheng et al. (Wed,) studied this question.