Using a training cohort of about 12,000 biopsies from the Cancer Genome Atlas which has been stained, fixed, and digitally enhanced, the CNN repeatedly analyzed normal renal biopsies and ccRCC core to accurately identify the presence of ccRCC. Ten cycles were carried out to test the reliability of the network. The CNN is able to generate 32 feature detectors, pooling layers, 6 fully connected hidden layer, and an output layer with sigmoid activation function. Using the Adam algorithm, the CNN was trained to test the validity and accuracy of properly diagnosing ccRCC in digitial core biopsy slides.
The CNN was able to accurately and rapidly analyze 2,480 slides in 1.5 seconds. Following 10 cycles of training data, the CNN had an accuracy rate of 99.98%. The area observed on these slides were about 25 µm2. The average time to process and analyze one specimen on a mid-range computer was 125 milliseconds.
Overall, CNNs are extrmely reliable and accurate in identifying ccRCC on renal core biopsies. Although the samples were small, the range and accuracy of CNNs were ample enough to successfully detect ccRCC. Future studies will focus on other histological subtypes of cancer, working with pathologists and radiologists to further develop CNNs, and assess paramters that may identify prognoses. Much like self-driving cars, it is able to carry out its main job, but there are still small mistakes that need to be perfected before it is used in a promising manner. Among these networks, other technological companies such as Microsoft have also developed their own networks to detect cancers, but none have reached the accuracy level of CNNs of this study.
Presented by: Michael Fenstermaker
Written by: Sherry Lu, Department of Urology, University of California-Irvine at the 2018 AUA Annual Meeting - May 18 - 21, 2018 – San Francisco, CA USA