Literature review summary of lymph node metastasis (LNM) prediction in CRC using deep learning, split by patient-level or individual LN assessment
| Patient-level LN assessment | |||||||
| Study details | Data summary | Model architecture | Key limitation | Sensitivity | Specificity | Accuracy | AUC |
| Bedrikovetski et al,16 2023, Australia | 1201 CT scans, CC, 500 scans with LN annotations | Shared ResNet-50 for segmentation and MLP classification | Non-expert annotations | 0.966 | 0.052 | 37.4% | 0.542 |
| Wan et al,18 2023, China | 610 MRI scans, RC, *limited to T1–2, tumour annotations | LN level 2D and 3D ResNet with transfer learning and averaging predictions from logistic regression | Not using cross-validation | 1.000 | 0.660 | 73.0% | 0.790 |
| Liu et al,21 2023, China | 282 CT scans, RC, Tumour annotations | ResNet-50, clinical and radiomics features, ML models and nomogram classifier | Test set of 57 | 0.955 | 0.857 | 89.5% | 0.942 |
| Xie et al,22 2023, China | 391 CT scans, CRC, LN annotations | LN and slice level ResNet-18 with an attention mechanism and logistic regression | No external validation | 0.734 | 74.7% | 0.768 | |
| Ding et al,17 2020, China | 545 MRI scans, RC, LN annotations | Faster R-CNN with transfer learning, clinical features, nomogram classifier and bounding box | 153/183 in the test set had LNM | 0.920 | |||
| Glaser et al,23 2020, Australia | 123 CT scans, CC, 595 individual LN annotations | DenseNet encoder-decoder with segmentation and MLP classifier | Test set of 23 patients | 0.860 | |||
| Individual LN assessment | |||||||
| Study details | Data summary | Model architecture | Key limitation | Sensitivity | Specificity | Accuracy | AUC |
| Ozaki et al,25 2023, Japan | 3547 LNs on MRI, RC, *Includes patients who had chemoradiotherapy | ResNet-18 with transfer learning, logistic regression | Used same pathology label for pretreatment/post-treatment | 0.963 | |||
| Li et al,24 2021, China | 129 MRI scans, RC | Inception-v3 with transfer learning, logistic regression | Only included 1–2 LNs per patient | 0.953 | 0.952 | 95.7% | 0.994 |
| Li et al,29 2021, China | 3364 MRI scans, CRC | LeNet/AlexNet with transfer learning, ML models and MLP | Used radiologist LNM diagnosis | 0.800 | 0.800 | 75.8% | 0.794 |
| Li et al,30 2021, China | 3364 MRI scans, CRC | AlexNet with transfer learning and skip connections, MLP | Used radiologist LNM diagnosis | 0.873 | 0.874 | 83.6% | 0.857 |
| Ding et al,32 2019, China | 414 MRI scans, RC | Faster R-CNN with transfer learning, MLP | All patients had LNM | ||||
| Li et al,28 2018, China | 619 LNs on MRI, CRC | Inception-v3 with transfer learning, MLP | Used radiologist LNM diagnosis | 94.4% | |||
| Lu et al,31 2018, China | 414 MRI scans, RC | Faster R-CNN with VVG16 and transfer learning, MLP and bounding box | All patients had LNM | 0.912 | |||
Includes study details, a description of the data with the number of patients or number of LNs, image modality (MRI/CT), cancer location (CRC or RC or CC) and unique inclusion criteria marked with an asterisk. Also, a model architecture summary, a key limitation and model performance metrics: sensitivity, specificity, accuracy and AUC to three significant figures. The best result for each metric is indicated in bold.
AUC, area under the curve; CC, colon cancer; CRC, colorectal cancer; RC, rectal cancer.