Room: Room 202
Purpose: To develop a convolutional neural network (CNN) that can directly estimate material densities based on multi-energy images acquired with photon counting detector (PCD) CT without going through an explicit conventional material decomposition process.
Methods: A CNN was constructed with three convolutional layers, two batch-normalization layers, and two rectified linear units. The sizes of the convolutional filters in 1st, 2nd, and 3rd convolutional layers were 3Ã—3Ã—2Ã—32, 3Ã—3Ã—32Ã—32, and 3Ã—3Ã—32Ã—2, respectively. We scanned a Gammex multi-energy CT phantom (Sun Nuclear. Corp) and a living swine (head and neck CT angiogram) on a research PCD-CT system (Siemens), using two energy bins ([30, 52] and [52, 120] keV). The phantom contained two hydroxyapatite (HA) (200 and 400 mg/cc), six Iodine (2, 5, 5, 5, 10, and 15 mg/cc), one adipose, and one solid-water inserts. Data augmentation strategies (e.g. random patches, random noise, and normalization) were used to reduce over-fitting in CNN training. The CNN was trained to predict the mass density of HA and iodine using a randomly selected 90% of the phantom data (the remaining data were used for validation). The trained CNN was tested using the animal scans for qualitative evaluation.
Results: The CNN technique accurately predicted the mass density of HA inserts (396.4Â±6.9 and 200.2Â±9.5 mg/cc) and Iodine inserts (1.72Â±0.45, 5.64Â±0.56, 5.53Â±0.79, 4.54Â±0.91, 10.8Â±0.58, and 15.5Â±0.54 mg/cc) in phantom images. The CNN technique also accurately decomposed bony structures and contrast-enhanced blood vessels in the animal images.
Conclusion: CNN technique can provide a convenient and accurate tool for material decomposition in PCD-CT.
Funding Support, Disclosures, and Conflict of Interest: No financial interests Dr. Cynthia H McCollough received research grants from Siemens Healthcare Other co-authors have nothing to disclose