Deep learning radiomics fusion model to predict visceral pleural invasion of clinical stage IA lung adenocarcinoma: a multicenter study

Abstract

Aim: To assess the predictive performance, risk stratification capabilities, and auxiliary diagnostic utility of radiomics, deep learning, and fusion models in identifying visceral pleural invasion (VPI) in lung adenocarcinoma.

Materials and methods: A total of 449 patients (female:male, 263:186; 59.8 ± 10.5 years) diagnosed with clinical IA stage lung adenocarcinoma (LAC) from two distinct hospitals were enrolled in the study and divided into a training cohort (n = 289) and an external test cohort (n = 160). The fusion models were constructed from the feature level and the decision level respectively. A comprehensive analysis was conducted to assess the prediction ability and prognostic value of radiomics, deep learning, and fusion models. The diagnostic performance of radiologists of varying seniority with and without the assistance of the optimal model was compared.

Results: The late fusion model demonstrated superior diagnostic performance (AUC = 0.812) compared to clinical (AUC = 0.650), radiomics (AUC = 0.710), deep learning (AUC = 0.770), and the early fusion models (AUC = 0.586) in the external test cohort. The multivariate Cox regression analysis showed that the VPI status predicted by the late fusion model were independently associated with patient disease-free survival (DFS) (p = 0.044). Furthermore, model assistance significantly improved radiologist performance, particularly for junior radiologists; the AUC increased by 0.133 (p < 0.001) reaching levels comparable to the senior radiologist without model assistance (AUC: 0.745 vs. 0.730, p = 0.790).

Conclusions: The proposed decision-level (late fusion) model significantly reducing the risk of overfitting and demonstrating excellent robustness in multicenter external validation, which can predict VPI status in LAC, aid in prognostic stratification, and assist radiologists in achieving higher diagnostic performance.

Keywords: Adenocarcinoma of lung; CT; Deep learning; Radiomics.

Fig. 1

The study design and pipeline

Fig. 2

The receiver operating characteristic curves (ROCs) of five distinct model in the trianing (left) and external test cohort (right)

Fig. 3

A comparative analysis of decision curves (depicted on the left) and calibration curves (depicted on the right) across various models is presented for the external test cohort. The X-axis denotes the threshold probability, whereas the Y-axis represents the net benefit achieved. The thin grey line represents the hypothetical scenario where all LACs patients are classified as high-risk, and the thin black line corresponds to the scenario where all patients are designated as low-risk. Calibration curves depict the calibration of each model in terms of the agreement between the predicted status of visceral pleura and observed outcomes of visceral pleura. The y-axis represents the actual visceral pleural invasion rate. The x-axis represents the predicted visceral pleural invasion. The diagonal dotted line represents a perfect prediction by an ideal model

Fig. 4

Kaplan–Meier survival curves evaluating the outcomes of clinical stage IA lung adenocarcinoma grouped according to their pathological VPI status (left), or their VPI status as predicted by the late fusion model (right). DFS, disease-free survival; VPI, visceral pleural invasion

Fig. 5

Performance comparison between radiologists with or without the late fusion model assistance in assessing the VPI status for the external test cohort