SUMMARY:
UCLA researchers in the Departments of Bioengineering and Electrical and Computer Engineering have developed a deep learning-based algorithm that can detect and quantify the breast cancer marker human epidermal growth factor receptor 2 (HER2) in microscopic images without the need for time-consuming immunohistochemical staining (IHC).
BACKGROUND:
One out of every eight women develops invasive breast cancer throughout their lives. The biomarker HER2 is often used to inform diagnostic decisions as well as guide cancer treatment plans. Typically, immunohistochemical staining is performed on a biopsy of the tissue to visualize HER2. This protocol requires a pathologist and generally takes one day to perform. The combination of expensive reagents, specialty equipment and expert hands-on time limits the number of samples that a pathologist can perform at once and contributes to the expense imposed by these tests. In addition, only a limited number of IHC stainings can be performed on the same sample, often forcing physicians to prioritize specific biomarkers and creating bottlenecks for additional diagnostic screening.
One exciting technology is the use of deep learning to visualize HER2 without the need for IHC. This works by imaging the natural fluorescence of the tissue known as autofluorescence. This data contains rich information about what biomarkers are present in the sample but are not interpretable to human eyes. Deep learning on the other hand can learn the subtle relationships in the autofluorescence images then predict and quantitate the specific biomarker levels in that sample. This has been shown to work well for a number of biomarkers but has yet to be applied to breast cancer diagnostics. There is a clear and pressing need for deep learning-based technologies that would allow the detection of HER2 in breast cancer to reduce dependence on the need to label samples via IHC.
INNOVATION:
Researchers at UCLA have developed a label-free virtual HER2 staining method that uses deep learning to quantify HER2 biomarker levels. The method has demonstrated comparable accuracy to IHC staining without the need for costly, laborious, and time-consuming IHC staining procedures. In a blind evaluation study, pathologists compared images that were labeled virtually to those labeled via standard IHC protocols. The quantitative studies indicated that the deep learning-based protocol displayed comparable level of staining qualities, providing virtually indistinguishable results. This method demonstrates an invaluable tool that would reduce diagnosticians’ dependence on IHC, and facilitate an efficient and cost-effective alternative to standard HER2 labeling protocols.
POTENTIAL APPLICATIONS:
• Label-free virtual staining of breast tissue
• Detection and quantification of breast cancer marker
• Potential expansion to other tissues
• Potential expansion to other biomarkers
ADVANTAGES:
• Rapid
• Cost-effective
• Reduces expert hands-on time
• No expensive reagents needed
• No expensive equipment needed
DEVELOPMENT-TO-DATE:
A successful demonstration of the technology has been performed via simulation and experimentally to show that the technology can accurately reconstruct HER2 IHC staining with only autofluorescence images.