MRI model predicts breast tumor shrinkage patterns

An MRI model that analyzes breast tumor microbiomes improved the prediction of tumor shrinkage patterns in a study published August 26 in Radiology.

A team led by Yuhong Huang, MD, from Southern Medical University in Guangzhou, China, reported high accuracy and overall performance for the model, which incorporates the intratumoral microbiome count, habitat radiomic features, and deep-learning features to predict tumor shrinkage patterns following neoadjuvant therapy.

“It also demonstrated consistent performance across molecular subtypes and clinical stages in the validation sets, highlighting its potential to help guide surgical planning following neoadjuvant therapy and to increase breast-conserving surgery success rates,” the Huang team wrote.

Women with breast cancer show different tumor shrinkage patterns after neoadjuvant therapy. Accurate prediction of these patterns is needed for guiding breast-conserving surgery. Previous reports suggest that the microbiome inside breast tumors influences response to treatment. The researchers, meanwhile, suggested that related imaging features could help improve the prediction of tumor shrinkage patterns.

Huang and colleagues developed and tested MRI models to predict these patterns, with MRI providing detailed information on tumor morphology and kinetics. They created five models that used 3D U-Net automated segmentation, habitat radiomic and/or deep-learning (ResNet-50) features, and histologic intratumoral microbiome data. The models included the following: pre-NAT habitat, mid-NAT habitat, pre-NAT ResNet-50, mid-NAT ResNet-50, and a fusion model.

The study included 2,249 women with breast cancer who had a median age of 49. The training set included 671 women, internal validation included 335 women, and external validation included 1,243 women.

Of the total women, 1,238 (55%) showed concentric tumor shrinkage. Tumors with concentric shrinkage had higher microbiome abundance (p < 0.001).

The 3D U-Net model achieved high Dice coefficient scores on both pre- and mid-neoadjuvant therapy MRI scans.

The fusion model, meanwhile, achieved the highest area under the receiver operating characteristic curve (AUC) in the internal validation (0.89, p < 0.05) and external validation sets (0.87, p < 0.001). The fusion model also achieved high AUC values across all molecular subtypes (AUC range, 0.85 to 0.91) and tumor stages (AUC range, 0.84 to 0.89).

The results show how specific radiomic and deep-learning features reflect microbiome abundance within breast tumors, the study authors highlighted. They added that the fusion model cuts the need for manual segmentation and feature extraction.

“This approach not only improves efficiency but also captures a broader range of tumor features through habitat radiomics,” they wrote.

The team called for prospective studies to confirm the model’s clinical impact and use in surgical workflows. The full study can be found here.