In the first blog post of this series on machine learning for the detection of breast cancer in medical imaging, we discussed how computers can learn how to interpret mammograms. The software developed by Dr. Becker and his colleagues from the University Hospital Zurich, Switzerland, reached a diagnostic accuracy comparable to that of radiologists, and it can be used as a second pair of eyes that quickly analyzes the medical images and instantly notifies the radiologists of its findings.
From mammography to ultrasonography
Let’s move to the second research presented that used machine learning to automate the detection of breast masses in ultrasonography. This technique uses sound waves to produce images, and we already mentioned that ultrasonography is mostly used when radiologists have identified suspicious features on mammograms. In addition, a study from 2001 showed that ultrasound associated with mammography increases the sensitivity to 77.5% compared to that of mammography alone (50%) in women with dense breasts and increased risk of breast cancer. However, breast lesions often appear to have discontinuous margins, which are difficult to trace. This is also why regular ultrasounds are frequently combined with Doppler ultrasound. This latter technique quantifies blood flow through blood vessels, and visualizes the presence and distribution of blood vessels associated with breast lesions. Therefore, it aids in determining whether the lesions are malignant or not. More precisely, malignant breast lesions are generally more vascular than benign ones. This means that if we want to automate the detection of breast cancer in ultrasound, we should probably combine regular and Doppler ultrasonography in order to maximize the diagnostic performance. That’s exactly what researchers at the University of Pennsylvania did.
Better automated diagnosis combining regular and Doppler ultrasounds
The research team first developed a machine learning algorithm in order to segment and classify lesions in regular ultrasound breast images. To do that, 8 major features to describe breast lesions in regular ultrasounds were extracted from the training set of images. Subsequently, they enrolled the computer in a second “medical training course” and taught it how to interpret Doppler ultrasounds. In the first case, with solely regular ultrasound images, the software diagnostic performance was rather good (AUC=0.88, sensitivity=76.1%, specificity=90%). Still, once Doppler features were added, the software interpretative skills heavily increased and almost reached perfection, reaching an AUC of nearly 0.95, sensitivity and specificity values above 91%. As you may remember, the human AUC values in detecting breast cancer in mammograms were between 0.83 and 0.94 and, therefore, the machine developed by the medical team in Pennsylvania is a reliable digital assistant that can help doctors in formulating more accurate breast cancer diagnoses.
The results obtained also indicate that, although we don’t really know how computers learn and how they interpret medical images, we can still identify some common patterns and similarities between trained physicians and trained machines. Stay tuned to find out how machine learning algorithms performed in the detection of breast cancer in histopathology images!