PET/CT-based deep learning for differentiating squamous cell carcinoma from adenocarcinoma in non-small cell lung cancer: comparison of multimodal fusion strategies.

Abstract

Objective: This study aimed to evaluate 18F-fluorodeoxyglucose (18F-FDG) PET/computed tomography (CT)-based deep learning for differentiating squamous cell carcinoma (SqCC) from adenocarcinoma (ADC) in non-small cell lung cancer (NSCLC) and to compare single-modality and multimodal fusion strategies.

Methods: This retrospective study included 220 patients with pathologically confirmed NSCLC (86 SqCC and 134 ADC) who underwent pretreatment 18F-FDG PET/CT. A 2.5D input strategy used the axial tumor slice with the largest mask area and its two adjacent slices. Four models were developed: PET-only, CT-only, PET/CT dual-branch fusion, and PET/CT six-channel early fusion. Performance was evaluated using five-fold stratified cross-validation with area under the receiver operating characteristic curve (AUC), accuracy, balanced accuracy, sensitivity, specificity, F1-score, and Matthews correlation coefficient (MCC).

Results: The PET-only model achieved higher mean AUC and MCC than the CT-only model. The PET/CT six-channel early fusion model achieved the highest mean AUC (0.766 ± 0.082), balanced accuracy (0.765 ± 0.071), specificity (0.882 ± 0.093), and MCC (0.530 ± 0.132). The dual-branch fusion model achieved the highest mean sensitivity (0.755 ± 0.144), although its mean AUC was lower than that of the PET-only model.

Conclusion: PET/CT six-channel early fusion showed potential for differentiating ADC from SqCC and may provide complementary information for noninvasive histological assessment. External validation is required before clinical application.