A study published by Academic Radiology has found that a nontraditional approach in training deep-learning algorithms resulted in higher specificity when categorizing lesions found on a breast MRI scan. The researchers, who work at Columbia University Medical Center, used a method of training deep-learning models using labels of image slices instead of manually made pixel-by-pixel notes to instruct a breast MRI lesion classification algorithm. The method is weak supervision, leading to high specificity and an area under the curve or AUC of 0.92.
Using this approach can vastly improve workflow since it creates large datasets more efficiently while paving the way for better breast MRI interpretation due to its high specificity. Here’s what you need to know about it:
How Weak Supervision Refines Workflow
Large, adequately annotated datasets are required to train neural networks properly. However, opting for training images labeled at the slice level allows technicians and surgeons to bypass annotating a region of interest pixel by pixel. This tedious work is often needed to create a large enough dataset, but weak supervision accelerates the workflow since it generates the required dataset more efficiently. It was also found to teach the network imaging features of normal anatomy and benign enhancement patterns of the breast, allowing for more precise imaging.
The researchers sought to identify weak supervision’s capacity to enable the algorithm to learn from the entire image rather than the delineated ROI. They aimed to evaluate the attainability of the method to improve breast MRI specificity, which is known for having a high false-positive rate.
How the Researchers Carried Out the Study
To uncover the effects of using weak supervision in MRI, the researchers gathered a dataset of 438 breast MRI studies, 307 of which came from their institution and 131 which came from other U.S. institutions. They sourced these datasets from the Investigation of Serial Studies to Predict Your Therapeutic Response with Imaging and Molecular Analysis or I-SPY TRIAL database.
Then, they combined almost 280,000 image slices from the studies into nearly 93,000 three-channel images. They used 85 percent for training and validating the convolutional neural network or CNN. The remaining 15 percent of the images were used to create a testing dataset, including 11,498 benign images and 1,531 malignant images.
The Test Set’s Performance
The researchers found that the algorithm attained 0.92 AUC, a specificity of 95.3 percent, a sensitivity of 74 percent, and an accuracy of 94.2 percent. These results concluded that it is viable to use a weakly supervised-based CNN to evaluate breast MRI images with high specificity. It offers numerous benefits, such as facilitating the compilation of larger datasets by removing the need to annotate them manually, pixel by pixel. It can also be used to lessen the subjective bias that sometimes occurs when human readers identify an ROI’s boundaries.
They also found that the network assesses the entire slice of MRI images instead of evaluating only the ROI, mimicking what occurs in real-life clinical practice. The researchers then concluded that weak supervision could result in better results for MRIs.
Conclusion
Thanks to the researchers’ work, facilities can address the high false-positive issue in breast MRIs using weak supervision. It also can improve results in other MRIs, introducing an innovative approach for clinics and practices that can make their workflow even more efficient.
DirectMed Parts is the leading supplier of MRI and CT parts and coils. From Siemens MRI parts to Philips MRI parts, we have what you need to make your machine perform more efficiently. Contact us today to find out how we can help you find the perfect MRI system that suits your needs!
