Clinical Scorecard: A Transparent Machine Learning Approach Using Ultrasound Radiomics for Preoperative Distinction Between Granulomatous Lobular Mastitis and Breast Cancer

At a Glance

Key Highlights

• Combined model integrating ultrasound radiomics and clinical factors achieved high diagnostic accuracy (AUC up to 0.935 in training cohort).
• Non-invasive, interpretable machine learning approach improves differentiation between GLM and breast cancer, aiding clinical decision-making.
• Shapley additive explanation (SHAP) analysis identified key imaging biomarkers strongly correlated with GLM prediction.

Guideline-Based Recommendations

Diagnosis

• Utilize ultrasound radiomics features combined with clinical predictors to improve preoperative differentiation of GLM from breast cancer.
• Consider machine learning models such as Extremely Randomized Trees, Light Gradient Boosting Machine, and Random Forest for radiomics analysis.
• Employ Shapley Additive Explanations (SHAP) to interpret model predictions and enhance transparency.

Management

• Tailor treatment strategies based on accurate differentiation: radical mastectomy for breast cancer versus pharmacological and/or surgical management for GLM.
• Use non-invasive imaging-based models to reduce reliance on invasive pathological sampling when appropriate.

Monitoring & Follow-up

• Apply decision curve analysis to evaluate clinical utility of predictive models in ongoing patient assessment.
• Monitor model performance metrics including AUC, accuracy, sensitivity, and specificity during validation.

Risks

• Be aware of potential misdiagnosis due to sonographic similarities between GLM and breast cancer without advanced radiomics analysis.
• Consider limitations of MRI including cost, contraindications, and contrast agent risks when selecting diagnostic modalities.

Patient & Prescribing Data

237 patients with pathological diagnoses of GLM or breast cancer undergoing preoperative ultrasound

Accurate preoperative differentiation supports appropriate treatment selection, potentially avoiding unnecessary radical surgery for GLM patients.

Clinical Best Practices

• Incorporate ultrasound radiomics with clinical risk factors to enhance diagnostic precision for breast lesions.
• Use interpretable machine learning models to facilitate clinician understanding and acceptance.
• Prefer non-invasive diagnostic tools to minimize patient trauma and improve workflow efficiency.
• Validate predictive models with independent cohorts to ensure generalizability and reliability.

References

• Granulomatous lobular mastitis clinical features and diagnosis
• Ultrasound imaging in breast disease differentiation
• Machine learning and radiomics in breast cancer diagnosis