Automated quantitative assessment of amorphous calcifications: Towards improved malignancy risk stratification

Area of Science:

• Medical Imaging
• Radiology
• Machine Learning in Healthcare

Background:

• Amorphous calcifications on mammograms present diagnostic challenges, leading to frequent biopsies with low cancer yield.
• Only 20% of biopsied amorphous calcifications are ultimately found to be cancerous.

Purpose of the Study:

• To develop and validate a quantitative approach for distinguishing benign from actionable amorphous calcifications.
• To reduce diagnostic uncertainty and potentially decrease the number of unnecessary breast biopsies.

Main Methods:

• Trained and validated a LightGBM classifier on 248 mammography images with biopsy-confirmed diagnoses.
• Extracted local (radiomic, region measurements) and global (distribution, expert-defined) features.
• Utilized k-means clustering for local features and concatenated them with global features.

Main Results:

• Achieved 100% sensitivity and 35% specificity on a test set of 60 images.
• The algorithm identified 25% of cases as benign, suggesting potential biopsy reduction.
• A decision threshold of 0.4 yielded a positive predictive value of 38%.

Conclusions:

• Quantitative analysis of mammograms can identify subtle features to differentiate calcification types.
• This approach shows promise in improving the accuracy of amorphous calcification assessment.
• Further development could lead to reduced invasive procedures for benign findings.