A Soft-Attention Guidance Stacked neural Network for neoadjuvant chemotherapy's pathological response diagnosis using breast dynamic contrast-enhanced MRI

Pathological complete response (pCR) to Neoadjuvant chemotherapy (NACT)is a significant clinical indicator for diagnosing patient outcomes and overall survival. However, the current diagnosis of pCR suffers from two limitations: firstly, it requires an invasive biopsy, intensifying the patient's pain and risk of infection. Secondly, current radiomics-based computer methods inefficiently utilize intrinsic characteristics of medical images. In this paper, we present a Soft -Attention Guidance Stacked neural Network (SAGS-Net) to address these limitations and predict the NACT responses from breast dynamic contrast-enhanced MRI (DCE-MRI). Particularly, We first design a self-adapting feature selection strategy to extract peritumoral features while excluding outside noises accurately. Then, SAGS-Net is built by stacking position-based spatial models that generate discriminative feature representation. Each stacked model inside has its semantic feature branch as a control gate for pCR-related feature selection. The semantic feature branch combined with the residual learning mechanism as a feature selector enhances valuable information and suppresses redundant information, simulating the process of radiologists making a clinical diagnosis based on domain knowledge. Finally, the incremental stacked network architecture assisted with the soft-attention strategy can gradually refine attention-aware features in complex DCE-MRI sequences. Experimental results based on real-world clinical datasets confirmed that the proposed SAGS-Net obtain superior performance with 93% AUC, and it provides a new way that leverage DCE-MRI sequences to predict NACT responses.

Automated summary

In this paper, a Soft-Attention Guidance Stacked neural Network (SAGS-Net) is proposed to predict Neoadjuvant chemotherapy (NACT) responses from breast dynamic contrast-enhanced MRI (DCE-MRI). SAGS-Net utilizes a self-adapting feature selection strategy to extract peritumoral features and employs stacked spatial models to generate discriminative feature representation. Experimental results show that SAGS-Net achieves superior performance on real-world clinical datasets.