CT-based deep learning radiomics predicts micropapillary/solid components in lung adenocarcinoma.

BACKGROUND

Micropapillary (MPP) and solid (SOL) components are high-grade histological patterns in lung adenocarcinoma (LUAD) and are associated with aggressive tumor biology and recurrence. Reliable preoperative identification using routine imaging remains challenging. This study aimed to develop and internally validate a computed tomography (CT)-based deep learning (DL) radiomics model for non-invasive prediction of MPP/SOL components in LUAD.

METHODS

We retrospectively enrolled 486 patients with surgically confirmed LUAD who underwent preoperative CT. Handcrafted intratumoral radiomics features, two-dimensional DL (DL2D) features, and clinical variables were used to develop radiomics, DL2D, clinical, and combined models. Model performance was evaluated in a hold-out test cohort using the area under the receiver operating characteristic curve (AUC), calibration analysis, and decision curve analysis (DCA).

RESULTS

The DL2D signature showed similar discrimination in the training and test cohorts, with AUCs of 0.849 [95% confidence interval (CI): 0.799-0.899] and 0.842 (95% CI: 0.765-0.919), respectively. The combined model achieved AUCs of 0.931 (95% CI: 0.903-0.959) and 0.865 (95% CI: 0.796-0.935), respectively. In the hold-out test cohort, the combined model outperformed the clinical and radiomics models but did not significantly improve discrimination compared with the standalone DL2D model (DeLong P=0.29). The decrease in AUC from the training cohort to the test cohort indicated possible optimism in the combined model. Calibration analysis and DCA suggested acceptable calibration and potential clinical utility within selected threshold ranges.

CONCLUSIONS

CT-based DL radiomics may support non-invasive prediction of MPP/SOL components in LUAD. Integrating DL, handcrafted radiomics, and clinical variables improved test-cohort discrimination compared with the clinical and radiomics models, but not compared with the standalone DL2D model. These findings warrant external and prospective validation and suggest that a simpler DL2D-only model may be sufficient for practical implementation if comparable performance is confirmed.