Deep CT model stratifies EGFR and TP53 mutation groups in lung adenocarcinoma

A deep-learning system that combined seven-channel computed tomography inputs with clinical and radiologist-assessed features identified five molecular groups in people with lung adenocarcinoma, outperforming conventional imaging-feature models in the study’s internal test set. The MedicalDINOv3 model achieved 70.1% accuracy, a macro-averaged area under the receiver operating characteristic curve of 0.879 and a macro-averaged F1 score of 0.690. Results from independent datasets, however, showed weaker and more uneven performance, particularly for tumors with EGFR exon 19 deletions, indicating that the approach is not yet ready to replace molecular testing.

The retrospective study began with 13,206 testing records and retained 7,229 traceable cases at Weifang People’s Hospital in China before applying pathology, imaging and timing criteria. The final internal cohort comprised 1,633 patients whose pretreatment chest CT and molecular status could be verified. Participants had a mean age of 63.75 ± 10.24 years, 49.1% were men and 72.4% had never smoked. Stage IV disease accounted for 58.7% of the cohort. The authors assigned patients to five categories: 731 with EGFR wild-type tumors, 295 with EGFR 19 deletions, 399 with EGFR L858R substitutions, 118 with EGFR 19 deletions plus TP53 mutations and 90 with EGFR L858R plus TP53 mutations.

TP53 status was not subdivided within the EGFR wild-type group, which therefore included both wild-type and mutant TP53 cases. The researchers said this asymmetric grouping introduced intra-group imaging phenotypic heterogeneity and blurred some boundaries between wild-type and mutant categories. It also means that the model’s wild-type discrimination represents separation of all EGFR wild-type cases from the four mutant groups, not identification of a strictly EGFR wild-type and TP53 wild-type population. Molecular status was based on routine single-gene EGFR assays or multigene panel testing, with formalin-fixed paraffin-embedded tissue used in most cases and cytological cell blocks used when tissue was unavailable or clinically required.

Only examinations meeting prespecified consistency and image-quality requirements were included. The baseline scan was the chest CT used for clinical assessment before treatment or tissue sampling, with scans obtained within seven days before planned treatment preferentially selected. When non-contrast and contrast-enhanced examinations were available in the same window, the contrast-enhanced study was preferred. The main acquisition parameters included 120 kV tube voltage, tube current of 200–500 mA, rotation time of 0.4–0.7 seconds, a 512 × 512 reconstruction matrix and B31f or B30f reconstruction kernels. Patients who had received antitumor treatment before imaging or molecular testing, or whose CT and pathology were separated by more than one month, were excluded.

For each lesion, radiologists selected key CT slices and generated regions of interest that were moderately expanded to preserve information about tumor margins and adjacent lung. The expansion ratio was set at 0.1 to balance margin preservation against background noise. Segment Anything Model 2 produced candidate masks in multimask mode, which two radiologists independently reviewed and revised while blinded to molecular labels. The pipeline retained the original grayscale image, lesion-only image, perilesional background, binary mask and a three-channel overlay, giving the encoder access to tumor, boundary and surrounding pulmonary information. Contrast-limited adaptive histogram equalization, with a clip limit of 2.0, and Z-score normalization were used to standardize intensity, while radiomic shape, first-order and texture features were extracted from the segmented volume for comparison models.

Quality-control results indicated that the imaging annotations were broadly consistent. Repeated semantic assessments of 60 randomly selected cases produced intraclass correlation coefficients above 0.90. In a separate segmentation analysis of 50 samples, the Dice similarity coefficient was 0.92, and all reported intraclass correlation coefficients again exceeded 0.90. Activation maps from three representative test cases showed high model responses within or near segmented tumors, although the investigators described this as qualitative evidence of lesion-focused attention rather than proof of a mutation-specific biological imaging mechanism.

MedicalDINOv3 used a pretrained DINO v3 Vision Transformer-Large backbone, a projection adapter for the seven-channel inputs and a medical feature adaptation module. Most backbone parameters remained frozen during training, although the final zero to six Transformer layers could be unfrozen. Inputs were resized to 224 × 224 and intensity-normalized before inference. Slice-level representations were aggregated for patients with multiple selected images, then combined with selected clinical and semantic CT features in a multilayer perceptron. The internal data were divided before model development into training, validation and test sets of 1,143, 327 and 163 patients, respectively. Preprocessing, feature selection, hyperparameter tuning and threshold setting were confined to the training and validation sets, while the test set was reserved for final reporting.

Training addressed class imbalance through inverse-frequency weighting and focal loss with a gamma of 1.0 for the deep model, and through training-set random oversampling for conventional machine-learning models. The validation and test sets retained their original distributions, and neither SMOTE nor generative augmentation was used. The study compared the fused MedicalDINOv3 system with a support vector machine using radiomics, a random forest using semantic CT features and a Qwen-based model using clinical information and textual semantic descriptions. All were evaluated with the same five-class framework. MedicalDINOv3 performed best numerically, compared with 52.7% accuracy, 0.792 macro-AUC and 0.524 macro-F1 for the support vector machine; 45.5% accuracy, 0.819 macro-AUC and 0.412 macro-F1 for the random forest; and 33.9% accuracy, 0.700 macro-AUC and 0.316 macro-F1 for the Qwen baseline.

Ablation experiments showed that the full combination was important to performance. A MedicalDINOv3 branch restricted to structured semantic features achieved 31.55% accuracy and a macro-AUC of 0.6305, while a branch receiving CT images alone achieved 55.61% accuracy and a macro-AUC of 0.6347. The CT-only branch performed better than the semantic-only branch, but its macro-AUC and macro-F1 remained below those of the radiomics-based support vector machine, indicating that image-only and fused performance should not be conflated. Class-specific internal AUCs for the fused model were 0.917 for EGFR wild type, 0.855 for EGFR 19 deletions, 0.894 for EGFR L858R, 0.854 for EGFR 19 deletion with TP53 co-mutation and 0.873 for EGFR L858R with TP53 co-mutation. Despite these values, the two common sensitizing EGFR classes retained overlap.

External testing produced more modest results. In 66 patients from the public TCIA, TCGA and CPTAC data, the model was evaluated only for binary EGFR status because detailed subtype and TP53 labels were unavailable. It achieved 67.0% accuracy and an AUC of 0.6666 for the division of 19 EGFR-mutant and 47 EGFR wild-type cases. A separate multicenter cohort of 116 Chinese patients supported five-class assessment: 71 were EGFR wild type, 18 had EGFR 19 deletions, 15 had EGFR L858R, seven had EGFR 19 deletions plus TP53 and five had EGFR L858R plus TP53. Model parameters were fixed and no additional training or fine-tuning was performed in either external dataset.

In the multicenter cohort, the model produced 75.5% raw accuracy, but balanced accuracy was 58.2%, macro-F1 was 0.574 and macro-AUC was 0.7243. The raw figure was inflated by the predominance of wild-type cases, which represented about 61% of that cohort. Subtype-specific external AUCs were 0.8342 for EGFR wild type, 0.6504 for EGFR 19 deletions, 0.7221 for EGFR L858R, 0.6586 for EGFR 19 deletion with TP53 co-mutation and 0.7563 for EGFR L858R with TP53 co-mutation. The study was constrained by its retrospective single-center development cohort, small numbers in the TP53 co-mutated groups, differences across centers and imaging conditions, absence of formal paired model comparisons and no confidence intervals. The authors concluded that CT-based prediction may eventually supplement molecular assessment, but larger, standardized multicenter cohorts and further validation are required before independent clinical use.

Subject of Research: CT-based deep-learning prediction of EGFR mutation subtypes and TP53 co-mutation status in lung adenocarcinoma

Article Title: CT-based prediction of EGFR mutation subtypes and TP53 co-mutation status in lung adenocarcinoma

Article References: Liu, H., Li, Z., Zhang, Y., Yu, G., Miao, Y., Teng, X., Xu, Y., Xu, W., Gao, C., Bai, F., & Li, L. (2026). CT-based prediction of EGFR mutation subtypes and TP53 co-mutation status in lung adenocarcinoma. Frontiers in Molecular Biosciences, 13, 1898841. https://doi.org/10.3389/fmolb.2026.1898841

Image Credits: AI Generated

DOI: 10.3389/fmolb.2026.1898841

Keywords: lung adenocarcinoma, EGFR mutation subtype, TP53 co-mutation, chest CT, deep learning, MedicalDINOv3, radiomics, molecular stratification, external validation, CT-based, prediction, EGFR