Combined model-driven and dual-cycle interactive strategy few-shot learning scheme for predicting breast cancer molecular subtypes based on DCE-MRI.

Background And Objective: Deep learning has achieved significant success in solving classification problems in the medical field, which has, in turn, facilitated the development of continuous and real-time health monitoring systems. Yet medical images have extremely low data regimes, unlike natural images with a large amount of data. Hence, to classify medical images with a few per-class data samples is a challenging problem. In this paper, we propose a novel few-shot learning scheme, which jointly embeds a model-driven mechanism and dual-cycle interactive strategy, to predict breast cancer molecular subtypes based on a few DCE-MRI samples.

Methods: We present a unique spatio-temporal recurrent network classifier (STRNC) that predicts breast cancer molecular subtypes by learning spatial correlations and temporal dependencies in DCE-MRI. Moreover, the proposed dual-cycle interactive strategy (DCIS) formulates N-way C-shot tasks by applying task-specific adaptation to the support set and conducting meta-level evaluation on the query set, thus improving the model's generalization to unseen tasks.

Results: Extensive experiments on public datasets have demonstrated that our proposed scheme achieves the best results, with a classification accuracy of 99.13 ±0.21 (%) for subtypes, and can accurately differentiate between individual molecular subtypes.

Conclusions: Overall, our proposed method is able to learn spatial correlation and temporal dependence in DCE-MRI and has the potential to guide clinical typing prediction.