A Bi-LSTM and multihead attention-based model incorporating radiomics signatures and radiological features for differentiating the main subtypes of lung adenocarcinoma

Background: The radiological features of computed tomography (CT) images and the sequence of radiomics signatures in continuous slices of lung CT lesions are helpful in identifying subtypes of lung adenocarcinoma. A model based on bidirectional long short-term memory (Bi-LSTM) and multihead attention can learn the rules of this sequence well.

Methods: In this study, 421 patients with 427 lesions confirmed as adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA) or invasive adenocarcinoma (IAC) were recruited from three hospitals. The radiomics signatures of the identified lesion regions in each CT image were extracted using 'PyRadiomics' software, and the corresponding radiological features were subsequently documented and collected. Then, the top 100 features were extracted by the minimum redundancy maximum relevance (mRMR) feature ranking method. A model based on the radiological and imaging features was established to classify the lesions using Bi-LSTM and multihead attention. The diagnostic performance of the model was measured by the area under the curve (AUC) of the receiver operating characteristic (ROC).

Results: The model combined radiological features and radiomics signatures. The AUCs of the model in the training, testing, and validation groups were 0.985, 0.94 and 0.981, respectively, and the accuracy was 0.92, 0.976 and 0.91, respectively. In addition, we trained two other models [convolutional neural network (CNN) + multihead attention, long short-term memory (LSTM) + multihead attention] and compared them using the testing dataset. The precision of the two models was 0.89 and 0.88, respectively, and the accuracy was 0.88 and 0.87, respectively.

Conclusions: Bi-LSTM and multihead attention based on radiomics signatures and radiological features provide a way to distinguish AIS, MIA, and IAC.