A comprehensive analysis of imaging features and clinical characteristics to differentiate malignant from non-malignant mammographic architectural distortion.

BACKGROUND

Mammographic architectural distortion (AD) is usually subtle and has variable presentations and causes, which poses a diagnostic challenge for breast radiologists and consequently a complex decision-making challenge for clinicians and patients. Presently, there is no reliable imaging standard to differentiate between malignant and benign ADs preoperatively. This study aimed to perform a comprehensive analysis of detailed mammographic and ultrasonographic features and clinical characteristics to enhance the diagnostic and differential efficacy for AD lesions. The findings have the potential to boost the diagnostic confidence of breast radiologists when encountering with AD lesions and could be instrumental in refining clinical management strategies for ADs.

METHODS

This retrospective study included consecutive female patients with ADs on screening or diagnostic mammography from January 6, 2015, to December 28, 2018. The patient's clinical data, mammographic and ultrasonographic or "second look" ultrasonographic findings, and pathological results were reviewed. The continuous variables were analyzed using the -test. The categorical variables were assessed using the Chi-square test or two-tailed Fisher's exact test. Logistic regression analyses were conducted to evaluate potential risk factors for pathologically proven malignant ADs. Machine learning model based on multimodal clinical and imaging features was constructed using R software.

RESULTS

Ultimately, 344 patients with 346 AD lesions were enrolled in the study (mean age: 47.40±10.07 years; range, 19-84 years). Of the ADs, 228 were malignant and 118 were non-malignant. Palpable AD on mammography was more likely to indicate malignancy than non-palpable AD (83.43% 49.15%, P<0.001). AD associated with other mammographic findings was more likely to be malignant than pure AD (73.58% 59.36%, P=0.005). Ultrasonography (US) correlates were observed in 345 of these 346 AD lesions. Among these US correlates, 63 (18.26%, 63/345) were detected by "second look" ultrasound. For the US correlates, the mammographic ADs that appeared as non-mass-like hypoechoic areas and masses on US were more likely to be malignant than those that appeared as other abnormalities (P<0.001). The sensitivity, specificity and diagnostic accuracy of the eXtreme Gradient Boosting (XGBoost) model based on clinical and comprehensive imaging features in differentiation of AD lesions in the validation set were 66.46%, 94.23% and 78.9%, respectively, and the AUC was 0.886 (95% confidence interval: 0.825-0.947).

CONCLUSIONS

The application of mammograms-guided "second-look" ultrasound could enhance the detection of US correlates, particularly non-mass-like features. The comprehensive analysis based on clinical and multimodal imaging features could be beneficial in improving the diagnostic and differential efficacy for AD lesions detected on mammography and instrumental in refining clinical management strategies for ADs.