TopoTxR: A topology-guided deep convolutional network for breast parenchyma learning on DCE-MRIs

Area of Science:

• Medical Imaging
• Computational Biology
• Radiology

Background:

• Breast parenchyma characterization in dynamic contrast-enhanced MRI (DCE-MRI) is complex.
• Existing quantitative methods like radiomics and deep learning struggle to explicitly quantify subtle parenchymal structures.
• Fibroglandular tissue complexity poses a challenge for accurate analysis.

Purpose of the Study:

• To develop a novel topological approach for explicit extraction of multi-scale topological structures in breast parenchyma.
• To integrate these topological structures into a deep-learning model using an attention mechanism.
• To enhance the prediction of neoadjuvant chemotherapy response by leveraging topological insights.

Main Methods:

• Proposed a topology-informed deep learning model, TopoTxR.
• Utilized multi-scale topological structure extraction to approximate breast parenchymal structures.
• Incorporated topological features into a deep learning prediction model via an attention mechanism.
• Validated the model using the VICTRE phantom breast dataset and public I-SPY 1 dataset.

Main Results:

• TopoTxR effectively approximates breast parenchymal structures using extracted topological features.
• The model demonstrates differential topological behavior in breast tissue related to chemotherapy response (pCR vs. non-pCR).
• TopoTxR achieved a 2.6% increase in accuracy and 4.6% AUC enhancement compared to state-of-the-art methods on I-SPY 1 and Rutgers datasets.

Conclusions:

• TopoTxR offers enhanced insights into breast tissues critical for disease pathophysiology and treatment response.
• The topological approach provides a more accurate approximation of complex breast parenchymal structures.
• This method shows significant potential for improving the prediction of neoadjuvant chemotherapy outcomes in breast cancer.