Table 2 Comparison of kidney image classification algorithm performance

TransMIL [16]

CAMELYON16/TCGA-NSCLC/TCGA-RCC

AUC: (CAMELYON16: 93.09%, TCGANSCLC: 96.03%, TCGA-RCC: 98.82%)

Using multiple instance learning (MIL) to explore morphological and spatial information in images

Mainly dealing with weakly supervised classification in whole-slice image (WSI)-based pathology diagnosis

CTransPath [84]

TCGA-RCC

AUC:99.1%

Self-computation of localized window attention using Swin-Transformer as a backbone model

Large amounts of unlabeled data are required

UGBC [85]

private dataset

ACC (glomerulus: 96.30%, Kidney: 96.60%)

Assigning image labels based on kidney-level classification using a high-throughput batch labeling scheme to exploit label noise immunity associated with deep neural networks (DNNs)

Dependence on the accuracy of label annotations

DenseNet201–Random Forest [86]

CT KIDNEY DATASET: Normal-Cyst-Tumor and Stone

ACC: 99.44% (cyst: 99.60%, kidney: 98.90%, tumor: 100%)

Feature extraction using deep migration learning model DenseNet-201-Random Forest

More resources are needed to train and use both models simultaneously

RCCGNet [89]

KMC-kidney dataset/BreakHis dataset

KMC-kidney (ACC: 90.14%, F1:89.06%)/BreakHis (ACC: 90.09%, F1: 88.90%)

RCCGNet contains a shared channel residual (SCR) block, which shares information between two different layers and complements each other's shared data

The model integration is complex