Purpose: To assess quantitative analysis and machine learning of Contrast Enhanced Mammography (CEM) for predicting biopsy outcome (cancerous versus benign lesions).
Materials and Methods: Under a HIPPA compliant IRB approved protocol, 161 consenting patients with BI-RADS 4A/4B/4C or 5 breast lesions detected with tomosynthesis and/or ultrasound underwent pre-biopsy CEM. CEM image acquisition order was dual low (LE)/high energy CC then MLO views with side ipsilateral to the index lesion acquired first for each view with digital subtraction (DES). Biopsy showed the entire cohort included 32 cancers (22 IDC, 3 ILC, 5 DCIS, 1 LCIS, 1 poorly differentiated carcinoma) and 129 benign lesions. Quantitative measures of background parenchymal enhancement (BPE) were calculated using an automatic pipeline on DES images excluding the index lesion(s). Two percentage BPE measures (%BPEb [over the entire breast] and %BPEf [over the dense tissue]) calculated with respect to a series of enhancement ratio values ranging from 0 to 2 (interval 0.1) were compared between cancer and benign groups, using Wilcoxon test with p≤0.05 considered significant. In addition, 107 radiomic features extracted from segmented lesions on the LE and DES images (individually and together) were modeled by machine learning to classify/predict the biopsy outcome of the lesions. Furthermore, mean lesion intensity changes over CC and MLO views were evaluated as a simple means of kinetic analysis to predict biopsy outcome.
Results: Both quantitative %BPEb and %BPEf for the breasts containing cancers vs. benign lesions were significantly different on both CC and MLO views, with %BPEf showing a more obvious and robust distinction. Radiomics models revealed best prediction performance on CC DES images (AUC=0.69), with a minimal improvement after adding on the MLO DES view (AUC=0.70; precision=79%), suggesting that the first view (CC) obtained was the best predictor of the biopsy outcome. This prediction is substantially higher than BI-RADS-based PPV3 of 20-40%. Radiomics models of LE images have significantly lower AUCs than DES images. The mean lesion intensity changes did not show a significant difference between cancerous and benign lesions.
Conclusion: Breasts containing cancers have higher BPE than benign breasts. Radiomics analysis of first acquired CEM CC view DES may offer substantial improvement in biopsy outcome prediction of malignancy vs. benignity for lesions rated by tomosynthesis and/or ultrasound as BIRADS ≥ 4A.
Clinical Relevance Statement: BPE quantified at CEM shows a similar role to the BPE derived at breast DCE-MRI; machine learning of breast lesion’s radiomics of CEM images may provide a much higher prediction on biopsy outcome.
.jpg "Chang Liu, BS photo")