The findings could help women avoid unnecessary surgery, wrote a team led by Dr. Richard Ha of Columbia University Medical Center.
"Given the widespread use of screening mammography, our application of machine learning has the potential to influence patient management such that patients predicted to have pure ADH lesions may undergo imaging surveillance rather than surgery," the group wrote.
ADH is often a precursor to invasive breast cancer but can be difficult to distinguish from ductal carcinoma in situ (DCIS), according to Ha and colleagues. Surgical excision is the current standard of care after ADH has been diagnosed by biopsy, but most women with the disease who undergo surgery do not end up having DCIS or invasive cancer. That's why it's important to find a way to determine which patients with ADH may be safely monitored rather than sent to surgery on diagnosis.
Ha's group sought to evaluate the feasibility of applying deep-learning algorithms based on convolutional neural networks (CNNs) to mammography data to distinguish ADH from DCIS.
For this study, Ha and colleagues used 298 images from 149 patients. Of these, 134 were from patients with ADH and 164 were from patients with DCIS. Eighty percent of the patients were assigned to a training-and-validation set for the network and 20% to a test set. The group used receiver operating characteristic (ROC) area under the curve (AUC) analysis to evaluate the CNN's performance.
The researchers found that the test set had an AUC value of 0.86, a sensitivity of 84.6%, and a specificity of 88.2% for distinguishing ADH from DCIS. The diagnostic accuracy was 86.7%.
"Our results indicate that it is feasible to apply a CNN algorithm to distinguish ADH from DCIS," Ha and colleagues wrote. "Further improvement of our algorithm with use of a larger dataset has the potential to result in the creation of a valuable tool that can be used in the clinical setting to treat patients less aggressively after ADH is diagnosed by biopsy."
Copyright © 2019 AuntMinnie.com