dcm_qa_cs_dl is a simple DICOM to NIfTI validator script and dataset. Some of these images are enhanced using compressed sensing or deep learning. Each manufacturer uses their own algorithms, and these samples allow users to determine the vendor specific private DICOM tags.
These images demonstrate Compressed Sensing (CS) and the Deep Learning based Deep Resolve Gain (DRG), Sharp (DRS) and Boost (DRB) filters. The images were acquired using a Siemens 1.5T Sola running XA51 and were provided by Paul S Morgan (University of Nottingham).
For Siemens, CS is currently only available for 3D turbo spin echo acquisitions. It is listed on the console as a iPAT (integrated Parallel Acquisition Techniques) method. So the iPAT options CS, GRAPPA or mSENSE are mutually exclusive.
Each series is saved as both enhanced and classic DICOM format.
Si_2_t1_space_sag_cs4: CSx4, TA 570.7sSi_3_t1_space_sag_NOcs: no CS, TA 160.4sSi_4_t2_tse_tra_p2_DR_off: GRAPPAx2 no DR, TA sSi_5_t2_tse_tra_p2_DR_on_BoostMidSharp: GRAPPAx2, DRB, DRS, TA 98.7sSi_6_t2_tse_tra_p2_DR_on_BoostHighSharp: GRAPPAx2, DRB, DRS, TA 98.7sSi_7_t2_tse_tra_p2_DR_on_Gain4_2_Sharp: GRAPPAx2 no DRG, DRS, TA 98.7sSi_8_t2_tse_tra_p2_DR_on_Gain2_5_Sharp: GRAPPAx2 no DRG, DRS, TA 98.7s
Detecting compressed sensing requires reading the private CSA Series Header Info (0021,1019). This will list a paradoxical combination where the total acceleration factor is greater than the product of in-plane and between-plane acceleration factors:
sPat.lAccelFactPE = 1
sPat.lAccelFact3D = 1
sPat.dTotalAccelFact = 4.0
which dcm2niix will translate to the BIDS field:
"CompressedSensingFactor": 4,
There are several forms of deep learning, using the Deep Resolve branding. Deep Resolve is only available for 2D acquisitions. I have seen three of these reported by XA51 in private tag (0021,1175): Deep Resolve Gain (DRG), Deep Resolve Boost (DRB), Deep Resolve Sharp (DRS). I have yet to see examples of Deep Resolve Swift Brain, so dcm2niix will not detect this.
For example, an image with Gain of 4 and Sharp of 2 will report:
(0021,1175) CS [ORIGINAL\PRIMARY\M\DRG\NORM\DRS\DIS2D]
(0021,1176) LO [ChannelMixing:ND=true_CMM=1_CDM=1\ACCAlgo:9\IterativeDenoising:RelStrength=0.850_MeanRelRisk=1.489\NormalizeAlgo:PreScan\EdgeEnhancement_2]
which dcm2niix will translate to BIDS (though note that the DR flags will appear in ImageType for classic DICOMs and ImageTypeText for enhanced DICOMs):
"ImageTypeText": ["ORIGINAL", "PRIMARY", "M", "DRG", "NORM", "DRS", "DIS2D"],
"DeepLearning": true,
"DeepLearningDetails": "ChannelMixing:ND=true_CMM=1_CDM=1\\ACCAlgo:9\\IterativeDenoising:RelStrength=0.850_MeanRelRisk=1.424\\NormalizeAlgo:PreScan\\EdgeEnhancement_2",
These images demonstrate HyperSense Compressed Sensing (CS) as well as AIR Recon DL Deep Learning reconstruction. These images simulate a GE 3T MR750 running MR30.1 and were provided by Jaemin Shin (GE HealthCare), see here for more details.
HyperSense (HS) must be used with 3D acquisitions (either gradient echo and spin echo) and ARC (GRAPPA) acceleration. It is not compatible with ASSET (SENSE). ARC can be 1x1 but it’s still required. The acceleration factor is for only the randomly sampled k-space locations. This factor doesn’t account for any ARC acceleration across ky-kz, partial fourier across ky-kz, or elliptical corner cutting typically used in HyperSENSE. Because HyperSense samples a radial (rather than square) k-space an agressive HS factor of 1.5 will reduce acquisition time by more than 2 times (100%*.7/1.5=47%). Note that HS is both in-plane and out-of-plane. Therefore, HS is not a good match for the existing BIDS ParallelReductionFactorInPlane tag.
GE_10_MPRAGE_P2S1H1.24: CSx1.24GE_13_Ax_T2_FLAIR_FS_DL_High: DL high (0.75)GE_14_Ax_T2_FLAIR_FS_DL_Off: DL off
HyperSense will be reported in a DICOM private tag (0043,10b7). For example, a HS of 1.24 will yield:
(0043,10b7) LO [1.24\1\10\0]
which dcm2niix will translate to the BIDS field:
"CompressedSensingFactor": 1.24,
AIR Recon DL will be reported in a DICOM private tag (0043,10CA). DL has three levels (low 0.3, medium 0.5 and high 0.75) and is available for both 2D and 3D acquisitions. For example, a DL of High will yield:
(0043,10ca) LO [0.75\High]
which dcm2niix will translate to the BIDS field:
"DeepLearningDetails": "0.75\\High",
Philips supports SmartSpeed artificial intelligence (AI) acceleration. The images were acquired using a Philips 3.0T Ingenia Elition X running software 11.0.0 and were provided by Paul S Morgan (University of Nottingham).
201_T1W_TSE_Sno_CSno_CSAIno: AI off (Enhanced DICOM)501_T1W_TSE_Sno_CSno_CSAI2: AI level 2 (Enhanced DICOM)
The presence of AI can be detected by looking for the flag AI in the private tag 2005,1110. However, the DICOM image does not appear to store the AI level.
(2005,1110) CS [CS_SENSE_AI ]
Philips brands Compressed Sensing as Compressed SENSE. Usage is mutually exclusive of the traditional SENSE acceleration. The acceleration factors for CSENSE and SENSE are similar (in the provided series 4, the scan would require 67 seconds without acceleration, and 34 seconds with either SENSE or CSENSE of x2). The images were acquired using a Philips 3.0T Ingenia Elition X running software 5.7.1 and were provided by Paul S Morgan (University of Nottingham).
Ph_401_T1W_FFE_Sno_CS2: CSx2 (Enhanced DICOM)Ph_501_T1W_FFE_Sno_CS4: CSx4 (Classic DICOM)
Philips reports compressed SENSE using public tags. It can be detected by the ParallelAcquisitionTechnique (0018,9078) tag reported CSENSE. Therefore, CS of 4 will yield:
(0018,9155) FD 1 # ParallelReductionFactorOutOfPlane
(0018,9168) FD 1 # ParallelReductionFactorSecondInPlane
(0018,9069) FD 4 # ParallelReductionFactorInPlane
(0018,9078) CS [CSENSE] # ParallelAcquisitionTechnique
Since the same tags are used as traditional acceleration dcm2niix will populate the existing BIDS fields:
"ParallelReductionFactorInPlane": 4,
"ParallelAcquisitionTechnique": "CSENSE",
Assuming that the executable dcm2niix is in your path, you should be able to simply run the script batch.sh from the terminal.