A deep learning model based on multiphase DCE-MRI for preoperative prediction of Ki-67 expression in breast cancer

This retrospective study was to develop and validate a deep learning model based on multi-phase Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) for non-invasive and accurate prediction of Ki-67 expression, a key proliferation biomarker critical for treatment decision-making and prognostic evaluation in breast cancer.

404 breast cancer patients who underwent preoperative DCE-MRI within 1 week of surgery were enrolled and randomly split into training (n = 282) and test (n = 122) sets in a 7:3 ratio. Multi-phase DCE-MRI sequences at 3.0T: pre-contrast phase, early phase (64 seconds), peak phase (128 seconds), and late phase (320 seconds) after contrast agent administration. DenseNet-121 was used to build four single-phase deep learning models (SP_DL1-SP_DL4). Their output probabilities (DL signatures) were combined using gradient boosting decision trees (GBDT) to create a multi-phase model (MP_GBDT). Clinical predictors were integrated with DL signatures to build a fused model (CMP_GBDT). Model interpretability was assessed using Grad-CAM and SHAP. Independent samples t-test or Mann-Whitney U test for continuous variables; χ ² test or Fisher's exact test for categorical variables; DeLong test for AUC comparisons. p ≤ 0.05 was considered statistically significant.

In the test set, single-phase DL models achieved AUCs of 0.712 (SP_DL1), 0.671 (SP_DL2), 0.761 (SP_DL3), and 0.664 (SP_DL4). The multi-phase DL model (MP_GBDT) achieved an AUC of 0.810, outperforming all single-phase models. The fused model (CMP_GBDT) reached a comparable AUC of 0.814, demonstrating no statistically significant improvement over MP_GBDT. SHAP identified SP_DL3 signature as the top contributor in both MP_GBDT and CMP_GBDT models.

The MP_GBDT model accurately and non-invasively predicted Ki-67 expression in breast cancer, with SP_DL3 signature being the main contributor.