Table 3 Performance comparison of kidney image algorithms for other applications
From: Application of visual transformer in renal image analysis
Algorithms | Datasets | Evaluation indicators/results | Main views and contributions | Usage |
|---|---|---|---|---|
VGG19 [87] | Private dataset | ACC: 98% | VGG19 uses a deeper network structure and a small convolutional kernel for improved feature extraction | Kidney cysts detection |
EANet [88] | CT KIDNEY DATASET: Normal-Cyst-Tumor and Stone | ACC: 83.65% | Introduction of attention mechanism, multi-scale feature fusion, efficient network design, cross-layer feature interaction | Kidney cyst classification |
ResNet50 [88] | CT KIDNEY DATASET: Normal-Cyst-Tumor and Stone | ACC: 87.92% | Having introduced Residual Block and Batch Normalization | Kidney cyst classification |
MD-BERT-LGBM [96] | private dataset | ACC: 78.12% AUC: 85.15% | The model integrates a bi-directional encoder representation of the Transformer with an optical gradient lifter, a multimodal data model | CKD disease prediction |
KidneyRegNet [90] | KiTS19/in-house datasets | KiTS19 (DSC: 96.88%, Sensitivity: 0.9711, Specificity: 0.9667)/in-house (DSC: 96.39%, Sensitivity: 0.9736, Specificity: 0.9560) | A new depth-alignment pipeline for free-breathing 3D CT and 2D U/S renal scans is proposed | Kidney alignment |
ChatGPT [23] | NA | NA | The core algorithm is the Transformer, which combines the Transformer model's self-attention mechanism with the language model's generative power | Nelson syndrome (NS) pathology report writing |
MulGT [100] | TCGA-KICA/TCGA-ESCA | KICA (Typing: AUC: 98.44%, ACC: 93.89%, F1: 93.89%, Staging: AUC: 80.22%, ACC: 74.98%, F1: 72.55%)/ESCA (Typing: AUC: 97.49%, ACC: 92.81%, F1: 92.74%, Staging: AUC: 71.48%, ACC: 66.63%, F1: 65.73%) | A domain knowledge-driven graph pooling module was designed to simulate diagnostic patterns for different analysis tasks | WSI task diagnostics |
Transformer [22] | DIVAT (Database of Kidney Transplantation Medical Records) | NA | For use in medical fields where continuous-time decision-making is required | Medical decision system |
Transformer [99] | Dataset of 56 renal biopsy WSIs in patients with DN | AUC: 97% | A multi-stage ESRD prediction framework based on the Transformer model | For encoding WSI (whole-slice images) and predicting future ESRDs |
Transformer [20] | Private dataset | F1: 96.3%, AUC: 98.9% | Predicting Kidney Transplant Function Using the Critical Mask Tensor of the Transformer Dot Product Attention Mechanism | Predicting kidney transplant function |
COVID-Net [21] | Private dataset | Survival prediction: ACC:93.55%, Kidney Injury Complications: ACC:88.05% | Proposing an interpretability-driven framework for building machine learning models to predict survival and kidney injury in patients with no coronary pneumonia from clinical and biochemical data | Predicting survival and kidney injury in patients with new crown pneumonia |
ExKidneyBERT [24] | Private dataset | OneQA (ACC: 83.3%) TwoQA (ACC: 95.8%) | Linguistic modeling of renal transplantation pathology reports | Renal pathology reports |