Implementation of machine learning into clinical breast MRI: Potential for objective and accurate decision-making in suspicious breast masses

Abstract

We investigated whether the integration of machine learning (ML) into MRI interpretation can provide accurate decision rules for the management of suspicious breast masses. A total of 173 consecutive patients with suspicious breast masses upon complementary assessment (BI-RADS IV/V: n = 100/76) received standardized breast MRI prior to histological verification. MRI findings were independently assessed by two observers (R1/R2: 5 years of experience/no experience in breast MRI) using six (semi-)quantitative imaging parameters. Interobserver variability was studied by ICC (intraclass correlation coefficient). A polynomial kernel function support vector machine was trained to differentiate between benign and malignant lesions based on the six imaging parameters and patient age. Ten-fold cross-validation was applied to prevent overfitting. Overall diagnostic accuracy and decision rules (rule-out criteria) to accurately exclude malignancy were evaluated. Results were integrated into a web application and published online. Malignant lesions were present in 107 patients (60.8%). Imaging features showed excellent interobserver variability (ICC: 0.81-0.98) with variable diagnostic accuracy (AUC: 0.65-0.82). Overall performance of the ML algorithm was high (AUC = 90.1%; BI-RADS IV: AUC = 91.6%). The ML algorithm provided decision rules to accurately rule-out malignancy with a false negative rate <1% in 31.3% of the BI-RADS IV cases. Thus, integration of ML into MRI interpretation can provide objective and accurate decision rules for the management of suspicious breast masses, and could help to reduce the number of potentially unnecessary biopsies.

Fig 1. Patient flow chart.

A database search revealed 254 patients having received breast MRI in our institute between 12/2013 and 06/2017 and classified as BI-RADS IV or V after complementary assessment by mammography, ultrasound or clinical examination. Exclusion criteria were: isolated non-mass-enhancements (n = 35), missing histological confirmation (n = 25), no suspicious lesion in MRI (n = 19), or an incomplete MRI protocol (n = 2). Applying the exclusion criteria resulted in n = 173 patients with n = 176 lesions in n = 176 breasts.

Fig 2. Assessment of T2w signal intensity (SI).

(A) T2w SI was assessed on fat-saturated T2-weighted sequences. The ROI mask (see main task) was copied to this sequence from the contrast-enhanced T1w sequence (B). The corresponding mean SI was normalized to the mean SI of the pectoralis major muscle. In this example, this resulted in a T2w SI of 3.0 (49.2/16.3).

Fig 3. Clinical cases: Breast MRI of three different patients with suspicious lesions.

Case 1: A 55-year-old woman presenting with a mass in her left breast, measuring 21 × 18 mm. ADC was 1015 × 10⁻⁶ mm²/s. There was a type-2 curve. T2w SI was 4.6. The SVM diagnosed malignancy (error rate / false positive rate: 2.9%). Histopathology: G3 NST. Case 2: A 50-year-old woman with a mass in her right breast, measuring 26 × 19 mm with an intermediate diffusion restriction (ADC 1300 × 10⁻⁶ mm²/s) and a type-2 curve. The T2w SI was 4.4. The SVM diagnosed malignancy (error rate / false positive rate: 15.9%). Histopathology: low-grade DCIS. Case 3: A 44-year-old woman with a mass in her right breast, measuring 15 × 15 mm without diffusion restriction (ADC 1545 × 10⁻⁶ mm²/s) and a type-1 contrast enhancement. The T2w SI was 12.7. The SVM excluded malignancy with an error rate / false negative rate of 2.8%. This diagnosis was correct and histopathology revealed a fibroadenoma.

Fig 4. Acquired parameters in benign and malignant lesions.

As demonstrated by boxplots (A to F) and the stacked column chart (G), all parameters showed significant potential for differential diagnosis (all, p ≤ 0.0007).

Fig 5. ROC analysis of the Machine Learning Algorithm.

The diagnostic performance in the entire study collective (full data set: AUC = 0.90; black line) and the subset of BI-RADS IV cases (AUC = 0.92; black dotted line) was very good, without significant difference (p = 0.67).

Fig 6. Exemplary output of the open-access internet application.

This application can be used to verify our results and to translate our findings into a clinical setting. It can be accessed at: http://bit.do/Breast-MRI In this screenshot, the results of case #1 of Fig 3 are demonstrated. The Machine Learning algorithm predicted malignancy with an error rate (false positive rate) of 2.9%, a PPV of 96.2% and a specificity of 97.1%. Histopathology revealed a G3 NST carcinoma.