The Value of Artificial Intelligence-Assisted Imaging in Identifying Diagnostic Markers of Sarcopenia in Patients with Cancer

Table 1

Summary of segmentation methods.

	Author (year)	Population	Mean age (year)	Localization	Neural network	Segmentation algorithm	Segmentation ground truth
1	Ackermans (2021) [19]	Cancer surgery cases, colorectal, ovarian, pancreatic cancers (training); polytrauma patients (testing)	Testing: 74	L3 muscle (L3M), intramuscular adipose tissue (IMAT), visceral adipose tissue (VAT), and subcutaneous adipose tissue (SAT)	DLNN	2D U-Net	Manual segmentation using software (TomoVision software “sliceOmatic”)
2	Borrelli (2021) [51]	Lymphoma (training); Prostate cancer (testing)	Training: 61; Testing: 67	L3	CNN	RECOMIA platform U-Net	Manual segmentation using cloud-based annotation tool (RECOMIA, http://www.recomia.org)
3	Castiglione (2021) [52]	Pediatric patients	0-18	Skeletal muscle area at the L3 level; 12-section or 18-section MIP images	CNN	U-Net	Manual segmentation
4	Amarasinghe (2021) [49]	Non-small-cell lung cancer	67	Skeletal muscle at the L3 vertebra	CNN+DL	2.5D U-Nets	Manual segmentation based on the Alberta protocol
5	Kim (2021) [58]	Gastric cancers receiving gastrectomy	60.4	L3	CNN	ResNet-18	Manual segmentation with software (Aquarius 3D workstation, TeraRecon)
6	Magudia (2021) [61]	Pancreatic adenocarcinoma	52	L3	CNN	DenseNet architecture model to predict spatial offset; U-Net architecture model for segment	Manual segmentation with software internal data set: sliceOmatic (TomoVision, Magog, Canada); external data set: OsiriX (Pixmeo, Bernex, Switzerland)
7	Koitka (2021) [59]	Individuals with abdominal CT scans (unknown patients)	Training: 62.6; Test: 65.6	Whole abdomen and not just on L3 slices	CNN	Multiresolution U-Net 3D	For annotation, the ITK Snap software (version 3.8.0) was used. Region segmentation was performed manually with a polygon tool
8	Hsu (2021) [57]	Pancreatic cancer	67	L3	CNN	ResNet-18 model for slice; 2D U-Net to segment	Manual annotated, expert labeled
9	Zopfs (2020) [16]	The Cancer Imaging Archive’s collection “CT Lymph Nodes” and the institutional picture archiving and communication system	62	Containing the abdomen and images above (cranial) and below (caudal) this region	DCNN	U-Net	Manual segmentation
10	Edwards (2020) [54]	Adult patients	18-75	L3	CNN	Supervised U-Net	Manual segmentation
11	Hemke (2020) [56]	200 subjects	49.9	Pelvic content	DCNN	U-Net	Manual segmentation using manual and semiautomated thresholding using the Osirix DICOM viewer (version 6.5.2, http://www.osirix-viewer.com/index.html)
12	Burns (2020) [47]	102 sequential patients	68	L1-L5	CNN	U-Net	Annotation utilizing ITK-SNAP software. Region segmentation was performed manually
13	Paris (2020) [48]	Critically ill, liver cirrhosis, pancreatic cancer, and clear cell renal cell carcinoma patients, renal and liver donors	Training/validation: 52.6; Test: 53.9	L3	DCNN	Adapt U-Net	Manually segmented by using SliceOmatic (TomoVision, Montreal, Canada, version 4.2, 4.3, and 5.0)
14	Blanc-Durand (2020) [46]	Unknown subjects	N/A	L3	DCNN	2D U-Net	Manually annotated using the public freeware 3DSlicer
15	Park (2020) [62]	Gastric cancer, pancreatic cancer, and sepsis and healthy individuals	Training: 56.1; Internal validation: 56.6; External validation: 61.1	L3	CNN	FCN-based	Semiautomated segmentation software (AsanJ-Morphometry) followed by manual correction
16	Barnard (2019) [50]	Older adults, who were current or former smokers	71.6	T12	CNN	U-Net	Manual segmentation using Mimics software (Materialise, Leuven, Belgium)
17	Graffy (2019) [55]	Asymptomatic adults	57.1	L3	CNN	U-Net	Manual segmentation
18	Dabiri (2019) [53]	Data from Cross Cancer Institute (CCI), University of Alberta, Canada	N/A	L3 and T4	CNN	FCN with VGG16	Manual segmentation using Slice-O-Matic V4.3 software (TomoVision, Montreal, Canada)
19	Lee (2017) [60]	Patients with lung cancer	63	L3	CNN	FCN of ImageNet pretrained model	Semiautomated threshold-based segmentation, followed by manual correction
20	Shephard (2015) [63]	N/A	N/A	N/A	N/A	N/A

L3M: L3 muscle; IMAT: intramuscular adipose tissue; VAT: visceral adipose tissue; SAT: subcutaneous adipose tissue; DLNN: deep learning neural network; CNN: convolutional neural network; MIP: maximum intensity projections; DL: deep learning; DCNN: deep convolutional neural network; N/A: not available; FCN: fully convolutional network.