Assessing breast cancer volume alterations post-neoadjuvant chemotherapy through DenseNet-201 deep learning analysis on DCE-MRI

Abstract

BackgroundChanges in tumor volume following neoadjuvant chemotherapy (NAC) are a crucial reference for determining surgical approaches in breast cancer, especially important for patients desiring breast conservation. MethodsBetween September 2015 and July 2022, 118 breast cancer cases from 109 patients were retrospectively gathered and randomly split into two cohorts: a training cohort of 83 cases and a test cohort of 35 cases. Deep learning models with DenseNet-201 architecture were constructed based on the peak enhanced phase of dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). Deep learning models were created to identify breast cancers that would experience tumor/node (T/N) stage downstaging or more than 90% reduction in MRI volume post-NAC. Their performance was compared with clinical models using receiver operator characteristic (ROC) curves for evaluation. The DeLong's test was used to determine significant differences in AUC among different models. ResultsIn the evaluation on a test cohort, a deep learning model showcased significantly better accuracy in predicting the downstaging of T/N stages in breast cancer with an AUC of 0.85. This performance markedly exceeded that of the traditional clinical model, which achieved an AUC of only 0.44. The deep learning model also excelled in predicting cases of breast cancer with more than 90% reduction in MRI volume, achieving an AUC of 0.89. This was superior to the clinical model's performance, which had an AUC of 0.61. The DeLong's test results showed that the predictive performance of the deep learning models was significantly better than that of the clinical models (P both <0.05). ConclusionThe deep learning model developed using DCE-MRI can accurately predict a tumor volume reduction of over 90% and downstaging of T/N stages in breast cancer patients after NAC, thereby potentially facilitating personalized treatment planning.