Additive value of texture analysis based on breast MRI for distinguishing between benign and malignant non-mass enhancement in premenopausal women

Yu Tan, Hui Mai, Zhiqing Huang, Li Zhang, Chengwei Li, Songxin Wu, Wen Tang, Yongxi Liu, Kuiming JiangBMC Medical Imaging 2021 · DOI

TL;DR

This study explores the use of texture analysis (TA) from breast MRI to improve the differentiation between benign and malignant non-mass enhancements (NMEs) in premenopausal women. By combining TA with traditional clinical and routine MRI characteristics (CRMC), the researchers developed a model that significantly enhances diagnostic accuracy, specificity, and sensitivity compared to using either method alone.

What Problem Does This Paper Address?

Distinguishing between benign and malignant non-mass enhancements (NMEs) on breast MRI is a significant diagnostic challenge. NMEs, unlike distinct masses, present as diffuse or irregular areas of enhancement, making them harder to characterize. This ambiguity can lead to unnecessary biopsies or delayed diagnosis of malignancy. While clinical and routine MRI characteristics (CRMC) offer some clues, their interpretation can be subjective and prone to inter-observer variability, especially in premenopausal women due to hormonal influences. Texture analysis (TA), a quantitative method that analyzes pixel intensity patterns, has shown promise in providing objective insights into tissue characteristics, but its additive value in a predictive model for NME diagnosis, particularly when combined with CRMC, remained unclear.

Methods

The study employed a retrospective design, analyzing MRI data from 147 premenopausal women with histologically confirmed NMEs. The data was split into training (n=102) and validation (n=45) sets.

  • Clinical and Routine MRI Characteristics (CRMC): Two experienced radiologists independently assessed routine MRI features based on BI-RADS guidelines, including lesion distribution (e.g., focal, linear, segmental) and internal enhancement patterns (e.g., homogeneous, heterogeneous). Time-intensity curve (TIC) patterns (persistent, plateau, washout) were also evaluated. Clinical data, specifically patient age, was also considered.
  • Texture Analysis (TA): Texture features were extracted from dynamic contrast-enhanced T1-weighted MRI sequences (2nd to 5th contrast phases) using the Mazda software. A region of interest (ROI) was manually delineated on the largest cross-section of the NME. A comprehensive set of texture features (over 300) were computed, categorized into six groups (RLM, ARM, Wavelet, GrM, Histogram, COM). A feature selection algorithm (FMC) was used to identify the top 30 texture features with the strongest ability to distinguish between benign and malignant lesions.
  • Combined Model: A multivariate logistic regression model was built, incorporating statistically significant factors identified from both CRMC and TA analyses. This combined model aimed to predict malignant NME.
  • Statistical Analysis: Mann-Whitney U tests and Pearson's chi-square tests were used for continuous and categorical variables, respectively. Univariate and multivariate logistic regression analyses were performed to identify significant predictors. Receiver operating characteristic (ROC) curves and the area under the curve (AUC) were used to assess the diagnostic performance of CRMC, TA, and the combined model.

Datasets

  • Training Dataset: 102 patients with NME.
  • Validation Dataset: 45 patients with NME.

The study did not introduce a new benchmark dataset but rather utilized existing patient data from the Guangdong Women and Children Hospital.

Key Insights and Ideas

  • TA's Complementary Role: While TA alone showed some diagnostic capability, its true value emerged when combined with CRMC. This suggests that TA provides information that is complementary to, rather than redundant with, traditional MRI assessments.
  • Importance of Dynamic Phases: Texture features extracted from the 2nd and 5th contrast phases were found to be particularly informative for distinguishing between benign and malignant NMEs. This highlights the dynamic nature of contrast enhancement and how its temporal evolution, captured by texture, can reveal underlying biological differences.
  • Reducing BPE Interference: The study's methodology included measures to minimize the impact of background parenchymal enhancement (BPE), a known confounder in premenopausal women. This focus on reducing BPE likely contributed to the improved specificity of the CRMC and combined models.
  • Age as a Significant Predictor: Patient age emerged as a significant independent factor in the combined model, indicating its continued relevance even with advanced imaging analysis.

Key Results

  • Combined Model Superiority: The combined model (CRMC + TA) achieved the highest diagnostic performance, with an AUC of 0.887 in the training set and 0.819 in the validation set.
  • CRMC Performance: CRMC alone yielded an AUC of 0.837, with a sensitivity of 80.6% and specificity of 72.5%.
  • TA Performance: TA alone showed a lower AUC of 0.74, with a sensitivity of 64.5% and specificity of 70%.
  • Improved Specificity with Combination: The combined model demonstrated improved specificity (80%) compared to CRMC alone (72.5%), suggesting it is better at correctly identifying benign lesions.
  • Significant Predictors: In the combined model, age, linear distribution, multiple-region distribution, S(5,5)Correlate, S(4,4)Correlate, and Perc.90% were identified as statistically significant independent factors.

Limitations and Future Work

  • Retrospective Design and Small Sample Size: The study acknowledges its retrospective nature and the use of a relatively small dataset, which could introduce bias. Future work should involve larger, prospective studies with diverse datasets and validation on different equipment.
  • Manual ROI Segmentation: The manual delineation of ROIs is prone to measurement errors. The authors suggest developing AI-driven tools for automated ROI segmentation to improve accuracy and reproducibility.
  • Generalizability: While the study focused on premenopausal women, further research could explore the applicability of this combined approach in other patient populations.

The authors suggest that the addition of texture analysis to routine MRI characteristics offers a promising, non-invasive quantitative approach for improving the diagnosis of NME, potentially reducing unnecessary biopsies and improving patient outcomes.

Summary updated

BibTeX
@article{tan2021additive,
  author = {Yu Tan and Hui Mai and Zhiqing Huang and Li Zhang and Chengwei Li and Songxin Wu and Huang Huang and Wen Tang and Yongxi Liu and Kuiming Jiang},
  title = {{Additive value of texture analysis based on breast MRI for distinguishing between benign and malignant non-mass enhancement in premenopausal women}},
  journal = {BMC Medical Imaging},
  year = {2021},
  doi = {10.1186/s12880-021-00571-x},
}
Figure