From this book "Deep Learning and Convolutional Neural Networks for Medical Image Computing"
About this time, Dr. Le Lu joined my group. An expert in computer vision, Le brought the passion and knowledge required to apply deep learning to the challenging problems we were investigating.
following lines is the 2 paragraphs before above
Another difficulty has been the time-consuming task of handcrafting of algorithms for CAD systems. Until recently, it was necessary to develop mathematical algorithms specifically tailored to a particular problem. For example, when I started to develop a CAD system for virtual bronchoscopy in 1997, there were no prior examples on which to build . My lab had to develop shape-based features to distinguish airway polyps from normal airways [9, 10]. When we extended the software to find polyps in the colon on CT colonography, it took about five years to perfect the software and annotate the images to the point where the software could undergo a robust evaluation on over 1000 cases . It took another five years for translation from the bench to the bedside . Other groups found it similarly time-consuming to develop, validate, and refine CAD systems for colonic polyp detection.
Most of our CADs used machine learning classifiers such as committees of sup- port vector machines, which appeared to be superior to other approaches including conventional neural networks, decision trees and linear discriminants [13–18]. About two years ago, I heard about “deep learning”, a new up-and-coming technology for machine learning . Deep learning was the name given to an improved type of neural network having more layers to permit higher levels of abstraction. Deep learn- ing was finding success at solving hard problems such as recognizing objects in real world images [20, 21]. An aspect of deep learning that caught my attention was its ability to learn the features from the training data. I had long been unhappy with CAD systems that required hand-chosen parameters and hand-crafted features designed for a particular application. I realized early on that such hand-tuning would not be reliable when the CAD software was applied to new data. In addition, the hand- tuning was very time-consuming and fragile. One could easily choose parameters that worked well on one dataset but would fail dramatically on a new dataset. Deep learning had the appeal of avoiding such hand-tuning.
Is "back then" or some other saying better than "About this time" to express when Dr. Le Lu joined the group?