Deep learning model for predicting spread through air spaces of lung adenocarcinoma based on transfer learning mechanism

Abstract

Background: Spread through air space (STAS) is a novel invasive pattern of lung adenocarcinoma (LUAD) associated with poor prognosis. Preoperative predicting of STAS helps choose an appropriate surgical and therapeutic strategy. This study aimed to develop and validate an STAS prediction model in LUAD based on deep learning algorithms.

Methods: A dataset of 290 patients with preoperative chest computed tomography (CT) images and confirmed STAS status was retrospectively selected. Optimal semantic features were selected by logistic regression. Image features were learned from cubic patches containing lung tumors and the area around the tumor within 5/10/15 mm extracted from CT scans. ResNet50 architecture was used to train deep learning models based on the transfer learning mechanism. The optimal semantic features are combined with the deep learning model to construct a hybrid model. Receiver operating characteristic (ROC) curves were used to evaluate the performance.

Results: Patients were randomly partitioned into a training set (70%, n=203) and a test set (30%, n=87). The International Association for the Study of Lung Cancer (IASLC) grade, maximum tumor diameter, tumor density, spiculated sign, vacuole sign, and peritumor obstructive inflammation were incorporated into the hybrid model as independent predictors. The STAS-HYBRID_t10 proved to be the optimal STAS prediction model with an area under the curve (AUC) value of 0.874 in the training set and 0.801 in the test set. The sensitivity, specificity, and accuracy of STAS-HYBRID_t10 were 0.659/0.526, 0.904/0.837, and 0.798/0.701 in the training set and test set, respectively.

Conclusions: The STAS-HYBRID_t10 has great potential for the preoperative prediction of STAS and may support decision-making for surgical and therapeutic planning in LUAD.

Keywords: ResNet50; Spread through air spaces (STAS); deep learning (DL); lung adenocarcinoma (LUAD); transfer learning.

Figure 1

Flowchart for the study cohort. CT, computed tomography; LUAD, lung adenocarcinoma; STAS, spread through air spaces.

Figure 2

Schematic diagram of ResNet50. ResNet50 started with conventional convolutional layers, batch normalization, ReLU functions, and max-pooling layers for initial feature extraction. The model then iterated through residual blocks to capture more higher-level features. Each residual block contained 3 convolutional layers. The grey arrow indicated residual connection. ResNet50 output the predicted probability of STAS-DL after the fully connected layer. In contrast, the semantic independent predictors were integrated into a feature combination process after the global average pooling layer, which then generated the predicted probability of STAS-HYBRID. CT, computed tomography; CONV, convolution; DL, deep learning; LUAD, lung adenocarcinoma; ROIs, regions of interest; ReLU, rectified linear unit; STAS, spread through air space.

Figure 3

ROC curves of STAS-MCS for STAS status prediction. (A) ROC curves for the training set of the STAS-MCS. The grey line indicates AUC values for resampling 1,000 times. (B) ROC curves for the test set of the STAS-MCS. AUC, area under the curve; MCS, morphological and clinicopathological signature; ROC, receiver operating characteristic; STAS, spread through air space.

Figure 4

Comparison of ROC curves between STAS-DL_t10 and STAS-HYBRID_t10. (A) ROC curves for the training set and test set of the STAS-DL_t10. (B) ROC curves for the training set and test set of the STAS-HYBRID_t10. AUC, area under the curve; DL, deep learning; ROC, receiver operating characteristic; STAS, spread through air space.

Figure 5

Representative case examples. (A) A solid lesion predicted to be STAS-positive by STAS-HYBRID_t10 and ultimately confirmed to be a true positive case. (B) A ground glass predominant nodule predicted as negative by STAS-MCS, but predicted as positive by STAS-DL_t10 and STAS-HYBRID_t10. (C) A case predicted to be negative by STAS-HYBRID_t10, but misclassified as STAS-positive by the STAS-MCS and STAS-DL_t10. (D) A false negative case. (E,F) Two false positive cases predicted by STAS-HYBRID_t10. DL, deep learning; MCS, morphological and clinicopathological signature; STAS, spread through air space.