A machine learning-based 18F-FDG PET/CT multi-modality fusion radiomics model to predict Mediastinal-Hilar lymph node metastasis in NSCLC: a multi-centre study.

AIM: To develop and validate a machine learning (ML) model based on positron emission tomography/computed tomography (PET/CT) multi-modality fusion radiomics to improve the prediction efficiency of mediastinal-hilar lymph node metastasis (LNM).

MATERIALS AND

METHODS

Eighty-eight non-small cell lung cancer (NSCLC) patients with 559 LNs from centre 1 were divided into training and internal validation cohorts (7:3 ratio), and 75 patients with 543 LNs from centre 2 were assigned as external validation cohorts. PET and CT images were fused by wavelet transform. Multi-modality fusion radiomics features from six images of lymph nodes were extracted. The multi-modality fusion radiomics (MFR), multi-modality fusion radiomics + metabolic parameters (MFRM), CT, PET and PET + CT models were developed based on the best one among the 11 ML algorithms. The receiver operating characteristic (ROC) curve and the Delong test were used to assess and compare the performance of the models.

RESULTS

The CatBoost algorithm was chosen, and the MFR, MFRM, CT, PET and PET + CT models were constructed. The MFR and MFRM models showed a high AUC for predicting LNM in centre 1 (AUC = 0.950 and 0.952) and centre 2 (AUC = 0.923 and 0.927), and there were significant differences in centre 2 (P=0.036). The diagnostic efficacy of MFR and MFRM models was significantly higher than CT, PET, PET + CT models and SUVmax≥3.5 (P<0.001). The MFRM prediction was statistically different from the MFR prediction in the hilar/interlobar zone.

CONCLUSION

Both the MFR and MFRM models based on multi-modality fusion radiomics showed great potential for non-invasively predicting mediastinal-hilar LNM in NSCLC.