Deep learning technology may help with breast positioning for mammography: Study
Japan: A recent study published in Scientific Reports has suggested that a deep convolutional neural network (DCNN) can be used to classify mammographic breast positioning to evaluate imaging accuracy.
Researchers found that the DCNN method had moderate accuracy for breast positioning classification and nipple profile and could help with breast positioning for mammography.
"The recognition of positioning criteria accuracy provides feedback to radiographers and can improve the accuracy of mammographic techniques," Haruyuki Watanabe, Gunma Prefectural College of Health Sciences, Maebashi, Japan, and colleagues wrote in their study.
Previous studies have shown inappropriate breast positioning as a common cause of mammographic imaging failure. Technologists must be trained in proper positioning, but such training can be complex and labour-intensive due to the subjective evaluation of mammography through visual inspection.
Artificial intelligence (AI) is being increasingly implemented by breast imaging departments in the screening process of breast cancer. DCNN-based learning, in particular, has shown the ability to distinguish benign from malignant breast lesions.
According to the study authors, there are no reports on DCNN use for verification of breast positioning.
To fill this knowledge gap, they proposed a DCNN classification for the validation and quality control of breast positioning criteria in mammography. Two main steps were designed for mammographic verification: automated detection of the positioning part and classification of three scales that decide the positioning quality through DCNNs. They collected 1631 mediolateral oblique mammographic views from an open database.
The authors used image processing to automatically detect the region of interest after acquiring labelled mammograms with three visual evaluation measures based on guidelines. The researchers then classified mammographic positioning accuracy using four representative DCNNs demonstrated in previous studies. These included four previously established DCNN frameworks: EfficientNet-B0, Inception-ResNet-v2, Xception, Inception-v3, and VGG-16.
The authors reported the following findings:
- The experimental results revealed that the DCNN model achieved the best positioning classification accuracy of 0.7836 in the inframammary fold using VGG16 and a classification accuracy of 0.7278 in the nipple profile using Xception.
- The breast positioning criteria could be assessed quantitatively by presenting the predicted value using the softmax function, which is the probability of determining positioning accuracy.
"The proposed method can be evaluated quantitatively without requiring an individual qualitative evaluation. It has the potential to improve the validation and quality control of breast positioning criteria in mammography," the team concluded.
They plan to explore a quality control system using DCNNs on other imaging positioning applications and ease the classification performance on the latest mammographic database.
Reference:
Watanabe, H., Hayashi, S., Kondo, Y., Matsuyama, E., Hayashi, N., Ogura, T., & Shimosegawa, M. (2023). Quality control system for mammographic breast positioning using deep learning. Scientific Reports, 13(1), 1-8. https://doi.org/10.1038/s41598-023-34380-9
