Table 1 Knee, hip and lumbar spine datasets and splits.

From: Region of interest-specific loss functions improve T₂ quantification with ultrafast T₂ mapping MRI sequences in knee, hip and lumbar spine

		Knee	Hip	Lumbar spine
Acquisition parameters	Scanner (s)	GE Discovery MR750w (GE Healthcare, Waukesha, WI), GE Discovery MR750 (GE Healthcare, Waukesha, WI)	GE Discovery MR750w (GE Healthcare, Waukesha, WI), GE Discovery MR750 (GE Healthcare, Waukesha, WI)	GE Discovery MR750w (GE Healthcare, Waukesha, WI), GE Signa PET/MR (GE Healthcare, Waukesha, WI)
	Coil(s)	8-channel T/R knee array (Invivo, Gainesville, FL)	32-channel cardiac array (Invivo, Gainesville, FL)	Geometry embracing method (GEM) posterior array (GE Healthcare, Aurora, OH)
	FOV	14 × 14 cm²	14 × 14 cm²	20 × 20 cm²
	acquisition matrix	256 × 128	256 × 128	256 × 128
	Slice thickness	4.0 mm	4.0 mm	8.0 mm
	Slices	22	28	12
	TEs	0 ms, 12.9 ms, 25.7 ms, 51.4 ms	0 ms, 10.4 ms, 20.8 ms, 41.7 ms	0 ms, 12.9 ms, 25.7 ms, 51.4 ms
	Readout BW	± 62.5 kHz	± 62.5 kHz	± 62.5 kHz
	Magnetization Recovery time	1.3 s	1.2 s	1.5 s
	ARC	2X	2X	None
	no phase wrap	None	2X k_y oversampling	None
	Other	64-view acquisition/T₂ preparation	64-view acquisition/T₂ preparation	64-view acquisition/T₂ preparation
Demographics information	Sex (M/F)	140/104	35/32	10/11
	Age	29.7 ± 12.9	48.9 ± 13.2	45.3 ± 14.7
	Weight	74.3 ± 12.7 kg	69.8 ± 12.4 kg	69.6 ± 11.0 kg
Training information details	Learning rate	0.001	0.001	0.001
	Batch size	1	1	1
	Number of batches used for training	R = 2: 40,390 R = 3: 36,351 R = 4: 36,351 R = 6: 40,390 R = 8: 60,585 R = 10: 36,351 R = 12: 40,390	R = 2: 39,030 R = 3: 28,622 R = 4: 26,020 R = 6: 39,030 R = 8: 36,428 R = 10: 26,020 R = 12: 39,030	R = 2: 1120 R = 3: 1344 R = 4: 1120 R = 6: 1120 R = 8: 2016 R = 10: 1344 R = 12: 3360
Training	Patients	144	39	13
	Scans	265	59	14
	Slices	5591	1533	112
Validation	Patients	50	15	4
	Scans	91	15	5
	Slices	1952	390	42
Test	Patients	50	13	4
	Scans	90	15	5
	Slices	1928	390	40
Total	Patients	244	67	21
	Scans	446	89	24
	Slices	9471	2313	194

MAPSS acquisition parameters for all datasets, with corresponding training, validation and test splits. ARC refers to Auto-calibrating Reconstruction for Cartesian Imaging⁷⁰. For hip acquisitions, no phase wrap was applied: k_y was oversampled by a factor of 2X, with space outside the prescribed y-FOV eliminated after reconstruction. In some cases, multiple acquisitions were taken per patient due to having multiple knees/hips scanned, or due to having follow-up scans for the same patient. Age and weight are reported mean ± 1 standard deviation (s.d.). Datasets were split into training, validation and test, ensuring all scans of a particular patient were only placed into one of the three datasets. Unless otherwise noted, all results are reported on the test set are described by this table; to ensure robustness of trained pipeline to data splits, additional versions were trained on 2 more splits detailed in Supplementary Tables S2 and S13, with results on those splits described in Supplementary Tables S14, S15, S16.

Back to article page

Table 1 Knee, hip and lumbar spine datasets and splits.

Search

Quick links