Kenji Suzuki PhD, Presenter: Royalties, General Electric Company Royalties, Hologic, Inc Royalties, AlgoMedica Royalties, MEDIAN Technologies Royalties, Riverain Technologies, LLC Royalties, Toshiba Corporation Royalties, Mitsubishi Corporation

Toru Higaki PhD, Abstract Co-Author: Nothing to Disclose

Wataru Fukumoto, Abstract Co-Author: Nothing to Disclose

Kazuo Awai MD, Abstract Co-Author: Research Grant, Toshiba Corporation Research Grant, Hitachi Ltd Research Grant, Bayer AG Research Consultant, DAIICHI SANKYO Group Research Grant, Eisai Co, Ltd

Although CT has been shown to be effective for screening lung cancer, current radiation dose in CT is still high for screening population. Our purpose was to develop a “virtual” high-dose CT technology to convert ULDCT to HDCT images with less noise or artifact.

We developed a supervised pixel-based machine-learning technique to convert ULDCT into HDCT images. We trained our technique with ULDCT (4mAs, 120kVp, 5mm slice thickness) and corresponding “teaching” HDCT (120mAs, 120kVp) of an anthropomorphic chest phantom (Kyoto Kagaku, Kyoto, Japan). Once trained, our technique does not require HDCT any more, and it provides “virtual” HDCT where noise and artifact are substantially reduced. To test our technique, we collected ULDCT (6.0±3.5mAs, 120kVp, 0.14±0.08mSv, 5mm slice thickness) of 12 patients on multiple vendor CT scanners (GE LightSpeed VCT; Toshiba Aquilion ONE). To determine a dose reduction rate of our technology, we acquired 6 CT scans of the anthropomorphic chest phantom at 6 different radiation doses (4, 10, 20, 40, 60 and 120mAs; 120kVp). Contrast-to-noise ratio (CNR) was used to evaluate the image quality of CT.

Our “virtual” HDCT technology reduced noise and streak artifacts in ULDCT (0.1mSv) substantially, while maintaining anatomic structures and pathologies such as vessels and nodules. With our technology, the average CNR of ULDCT images was improved by 14.3±1.9dB (from -16.1±4.3 to -1.8±3.7dB) (two-tailed t-test; P<.05). This 14.3 dB CNR improvement was equivalent to a radiation dose reduction rate of 0.1 in our phantom study. The processing time for each case was 48 sec on a PC (AMD Athlon, 3.0GHz).

Our technology converted ULDCT to virtual HDCT where noise and streak artifacts were reduced substantially, and it can potentially reduce radiation dose by 90% in CT.

Substantial reduction of radiation dose in CT with our technology would be beneficial to screening population. Very short processing time is an advantage of our technology over iterative reconstruction.

Suzuki, K, Higaki, T, Fukumoto, W, Awai, K, “Virtual” High-dose CT: Converting Ultra-low-dose (ULD) to Higher-dose (HD) CT by Means of Supervised Pixel-based Machine-learning Technique.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14045679.html