Synopsis

Keywords: Breast, Diffusion/other diffusion imaging techniques

Diffusion weighted imaging (DWI) holds great potential in improving specificity of findings detected on contrast enhanced breast MRI. Restriction spectrum imaging (RSI), an advanced diffusion imaging model, has potential in discriminating between malignant and fibroglandular breast tissue. In this abstract, we evaluate RSI’s performance, combined with a random forest model, in differentiating lesions requiring biopsies from lesions that do not. The model showed significant preliminary results.

Introduction

Breast cancer MRI screening protocols include dynamic contrast-enhanced (DCE) imaging for lesion detection and assessment. DCE has variable specificity (61-97%)¹, requiring biopsies for final diagnosis. Diffusion-weighted imaging (DWI) is now often part of cancer screening protocols, aiming at improving specificity. Minimizing unnecessary biopsies is a challenge with both clinical and economic implications. DWI is widely used, but more advanced models such as restriction spectrum imaging (RSI) aim to separate signal components in different degrees of diffusion (free, hindered and restricted). An RSI breast-specific model was developed on cancers greater than 2cm in size², and has shown potential in distinguishing cancers from healthy tissue³, but also in discriminating cancers from benign lesions⁴. In this study, we explore and evaluate the performance of the RSI outputs with a random forest model as a potential breast lesion predictor on a prospective study, performed on a breast cancer screening cohort.

Methods

Women were recruited in two cohorts from a screening population– after screening mammography interpreted as BIRADS-0 with subsequent diagnostic work-up leading to the recommendation for biopsy or undergoing routine screening MRI due to high-risk status. A total of 187 women were prospectively scanned, of which 72 had at least either one malignant or one benign lesion and corresponding RSI data. A total of 89 lesions were identified. Lesions were then separated as 41 average-risk lesions (ARL), 13 high-risk lesions (HRL) and 35 cancers (CL). All lesions were assessed via core needle biopsy.
Data were collected using a 3.0T wide-bore (Discovery MR750w, GE Healthcare, USA) and an 8-ch Sentinel breast coil. Reduced field-of-view (FOV) DW-MRI and DCE were collected. DW-MRI data were processed using Matlab (R2017a, Mathworks, USA) and were distortion corrected using the reverse polarity gradient (RPG) method⁵, followed by noise and eddy current corrections. Finally, datasets were normalized by the 98th percentile of the b= 0 s/mm² images. Pre-processed data of all patients were fit to the breast RSI model2 to generate signal contribution maps of components C₁, C₂, C₃:

$$ S(b) = C_1 + C_2e^{-b×1.4×10^{-3}} + C_3e^{-b×10.3×10^{-3}} $$

The outputs are hypothesized to correspond to cancer and fatty tissue (C₁), fibroglandular tissue (C₂), and vascular flow (C₃), respectively. The square root of C₁C₂ was calculated, as well as the ratio $$$ \frac{C_1C_2}{C_3}$$$ (and divided 10-fold to better account for scaling) from Ci values in each ROI. All lesions were 3D-delineated using ITK-SNAP⁶ under the supervision of a radiologist, based on suspicious findings detected on DCE-imaging corresponding to biopsied tissue. Control ROIs were drawn in healthy fibroglandular tissue, while avoiding any breast lesion, once for each patient that had at least one biopsied lesion.
A random forest model was fit to the data, with the outcome as the type of lesion: control (healthy), ARL, HRL, and CL, and the predictors of interest include C₁, C₂, C₃, square root of C₁C₂, and $$$ \frac{C_1C_2}{10C_3}$$$ . The control and ARL types were combined into one type as no biopsy is required for both types, while HRL and CL were combined as they both require biopsy and close follow-ups and treatment/surgery.
The data of lesions were randomly split into a training set (90% of lesions) and a set-apart test set (10% of lesions). Then a random forest model was fit to the training set only without using data from the test set. Parameters in the random forest model were determined using a 5-fold cross validation. The model trained with only the training set was then applied to the test set.

Results

The prediction with regrouped population (control + ARL vs HRL + CL) accuracy is 85% with 95% CI: (79%, 91%) in the training set and 0.76 with 95% CI: (50%, 93%) in the test set. Classification results are shown in confusion matrices in Table 1.

Discussion

In this study, we propose a random forest model for breast lesion classification. Preliminary results have demonstrated potential in improving diagnostic specificity, especially when regrouping lesion in two groups “no biopsy required” (control + ARL) versus “biopsy required” (HRL + CL). Minimizing unnecessary biopsies is crucial for patient care, and a high-performing model would help reduce unnecessary procedures for patient and health care expenses.
Next steps to improve the model will start with including more lesions in the dataset, especially high-risk lesions as there are fewer (only 13 compared to 41 ARL and 35 CL). With optimal sample size, a more comprehensive analysis will be performed, including optimizing data partitioning for model training.
Finally, the model would benefit from better, high-resolution DWI acquisitions. Parallel-imaging techniques, such as multi-slice excitation, could help improve through-plane resolution and provide better datasets for model training.

Acknowledgements

No acknowledgement found.