UCSF

Invasive lobular carcinoma (ILC) of the breast is the second most common histologic subtype of breast cancer. A majority of ILCs are estrogen receptor-positive (ER+) with a diffuse growth pattern that is difficult to detect. Patients with ILC often present at a clinically advanced stage with a low recurrence risk score. The combination of “molecularly low risk” and “clinically high risk” attributed the unique diagnostic and treatment challenges of this type of breast cancer. Dedicated breast positron emission tomography with [18F]fluoroestradiol (FES-dbPET) with high sensitivity and spatial-resolution is a new functional imaging approach to characterize ER+ breast cancers. In this observational study, we hypothesized that FES-dbPET imaging followed by a radiomic-based analysis of the primary tumor might aid in-depth characterization of ILC. Methods: Patients with biopsy-confirmed locally advanced ILC were imaged with dbPET using 5 mCi of FES before treatment. The primary tumor 3D volume was segmented from the ipsilateral breast. The segmentation of the whole contralateral breast volume was also obtained. Standardized uptake values (SUVs), background uptake values (BPUs), and radiomic features were computed. The top 9 radiomic features were selected for further analysis using the “Maximum Relevance – Minimum Redundancy” (mRMR) machine learning algorithm. Spearman rank correlation and Wilcoxon rank-sum test were performed to assess the relationship between imaging measurements and tumor characteristics, such as size and growth. All statistical analysis was performed using Python v 3.9 with Pandas v. 0.23.0, Numpy v. 1.21.0, Pingouin v. 0.4.0, Scipy v. 1.7.0, MatPlotLib v. 2.2.2, and Seaborn v. 0.11.1 packages to execute calculations and construct figures. P-values with α= 0.05 were calculated and reported for all measurements to establish the level of statistical significance. Results: A cohort of 15 ILC patients was included in this analysis. A total of 107 radiomic features were analyzed. A total of 12 radiomic features showed a statistically significant correlation with MRI tumor size, while only one radiomic feature correlated with Ki67 (p-value < 0.05). The tumor-background ratio (TBR) showed weak and insignificant correlation trends with tumor characteristics. No other significant correlations were found. Conclusion: This study demonstrated a whole-tumor methodology to characterize the early stage primary ILC, offering more in-depth information relating imaging features to tumor characteristics. Shape (4) and intensity (2) features, as well as non-uniformity (6) and gray-level zone emphasis (1) features from the textural analysis exhibited promising trends in characterizing ILC with respect to MRI tumor size and Ki67. Due sample size limitations, a larger cohort study is needed to verify the initial findings.