Integrating central-slice anatomy and MRF-mapped radiomics dynamics for lung adenocarcinoma subtyping.

Background: Lung adenocarcinoma is a leading subtype of lung cancer, and accurate subclassification is critical for guiding clinical management. However, existing computed tomography (CT)-based approaches often fail to capture how lesion characteristics evolve across adjacent slices. This study aims to address this gap by integrating central-slice anatomy with dynamic radiomics information. Methods: In this multi-center retrospective study, CT lesions were allocated into training, internal testing, and external validation cohorts. For each lesion, a 64×64 patch from the automatically selected central slice (IMG1) and a multi-layer radiomic sequence derived from PyRadiomics descriptors were extracted. After feature filtering and least absolute shrinkage and selection operator (LASSO) selection, 14 key descriptors were arranged slice-by-slice and modeled with a BiLSTM-Attention network. The resulting temporal representation was transformed into a two-dimensional Markov Random Field (MRF) map and fused with IMG1 to form a dual-channel input for the proposed dynamic radiomics fusion network (DRFN). DRFN was compared with two baselines: a convolutional neural network (CNN) using only IMG1 and a BiLSTM-Attention model using only radiomics sequences. Performance was assessed using area under the curve (AUC), accuracy, precision, recall, and F1-score, and interpretability was explored with class activation mapping (CAM). Results: On an independent test cohort, DRFN achieved per-class AUCs of 0.97 (minimally invasive adenocarcinoma), 0.99 (adenocarcinoma in situ), and 0.95 (invasive adenocarcinoma). Grad-CAM heatmaps confirmed that DRFN consistently attended to lesion cores, spiculated margins, and adjacent vascular structures-mirroring radiologists' diagnostic reasoning. Compared to a single-channel CNN (AUCs: 0.77-0.88) and a BiLSTM-Attention-only model (AUCs: 0.88-0.94), DRFN demonstrated superior sensitivity, specificity, and generalizability. Conclusions: By fusing static anatomical information with dynamic radiomics evolution maps, DRFN offers both high classification accuracy and transparent interpretability. Our framework thus holds promise as an intelligent diagnostic aid for lung adenocarcinoma subtyping in clinical practice.