University of Floridas AI finds early lung cancer with 97% sensitivity

Lung cancer is the leading cause of cancer death among men and women worldwide, according to the American Cancer Society. Each year, more people, about 154,000, die of lung cancer than from colon, breast, and prostate cancers combined, and the lifetime risk of developing lung cancer is as high as 1 in 15.
 
Successful patient outcomes depend on early detection, of the half of new patients diagnosed after lung cancer has spread, only 4 percent will live for five years. Fortunately, advances in artificial intelligence (AI) could make it easier for clinicians to spot signs of tumor growth more accurately than with eyes alone. 
 
A paper recently published on the preprint server Arxiv.org (“Single-Shot Single-Scale Lung Nodule Detection“) details a deep learning method for lung detection. Using a convolutional neural network (CNN) — a layered machine learning model that mimics the behavior of neurons in the human brain — researchers at the University of Florida’s Center for Research in Computer Vision (CRCV) were able to identify small nodules of lung cancer with 95 to 97 percent sensitivity.
 
Their work builds on that of NYU researchers in September, who retrained Google’s Inception v3, an open source convolutional neural network architected for object identification, to detect certain forms of lung cancers with 97 percent accuracy.
 
“Our approach uses a single feed-forward pass of a single network for detection and provides better performance when compared to the current literature,” the team wrote. “We used publically available … [scans] and showed that the proposed method outperforms the current literature both in terms of efficiency and accuracy … To the best of our knowledge, this is the first study to perform lung nodule detection in one step.”
 
The researchers’ system, dubbed S4ND, divided input data — computerized tomography (CT) — into a grid of cells, and performed classification for all cells simultaneously. Its 36 convolution layers (comprising five “densely connected” blocks of six layers) took into account contextual information from the entire scan to predict the presence of nodules in a cell, some of which were less than 3mm in size.
 
The researchers trained S4ND on a single Nvidia Titan XP GPU workstation with 64GB of RAM, feeding it 888 CT scans from the Luna dataset — a publicly available dataset of CT lung cancer scans annotated by human radiologists — sampled so that the nodes appeared in random locations (to avoid bias). They tested its accuracy by putting those scans through again, but this time shifted in four directions by 32 pixels.
 
The result? The S4ND was demonstrably better at handling the variation in textures, shapes, and position of nodules than traditional computer-aided detection systems, the researchers wrote, and had an easier time reconciling discrepancies between large search spaces (i.e., the entire lung) and the comparatively small nodes.
 
“We experimentally validate[d] the proposed network’s performance … on publicly available LUNA data set, with extensive comparison with the natural object detector networks as well as the state of the art lung nodule detection methods,” the researchers wrote. “A promising future direction will be to combine diagnosis stage with the detection.”