Table 3.
Summary of the application of AI in the renal ultrasound for diagnosis of kidney disease.
| Task | Algorithms | Data source | Size | Results | Ref. |
|---|---|---|---|---|---|
| Classification and diagnosis renal masses | SVM, logistic regression, naïve Bayes and quadratic discriminant analysis | clinical medical records of single center | 10 AMLs; 42 RCCs | ACC=94% | (37) |
| Radiomics nomogram based on ultrasound | clinical medical records of single center | 600 masses | AUC=0.91 ACC=90% |
(38) | |
| Quantitative texture information combined with tumor-to-cortex echo intensity ratio and tumor size | hyperechoic renal mass <5 cm in size | 105 AMLs; 25 RCCs | AUC=0.95 | (39) | |
| Using EffecientNet-b3 to extract features from B-mode and CEUS images, and adaptive weights are learned to fuse the features | B-mode and CEUS-mode images from two centers | 9794 B-mode and CEUS-mode images | ACC=80% Sen=80% Spe=79% AUC=0.88 |
(40) | |
| An ensemble of deep neural networks (ResNet-101, ShuffleNet, and MobileNet-v2) based on transfer learning | images from available standard datasets and radiologists | 4940 images | ACC=96% | (41) | |
| Diagnosis of CAKUT | MIL combined with transfer learning | sagittal and transvers view from the first renal ultrasound scans after birth were used | 86 patients with CAKUT; 96 controls |
AUC=0.97 ACC=94% |
(36) |
| a deep MIL method based on graph convolutional networks, instance-level and bag-level supervision | clinical medical records of single center | 120 CAKUT patients with 2687 images; 105 controls with 2246 images | ACC=85% Sen=86% Spe=84% |
(42) | |
| SVM classifiers integrating texture and deep transfer learning image features | clinical medical records of single center | 50 CAKUT patients; 50 controls |
AUC=0.92 ACC=87% |
(43) | |
| Diagnosis and grading of hydronephrosis | An encoder-decoder framework based on U-net | coronal and transverse view of the renal ultrasound scans | Labeled dataset: 1850 images; Graded dataset: 1407 images |
ACC=89% | (44) |
| An Attention-Unet which consists of four convolution blocks | clinical medical records of single center | 506 patients with hydronephrosis; 193 controls | Dice=0.83 Sen=90% Spe=80% |
(45) | |
| Keras neural network consists of five convolutional layers, a fully connected layer of 400 units, and a final output layer | sagittal view of the renal ultrasound scans | 2420 images | ACC=78% | (46) |
CAKUT, congenital abnormalities of the kidney and urinary tract; SVM, support vector machine; CEUS, contrast-enhanced ultrasound; MIL, multi-instance learning; CNN, convolutional neural network; AML, angiomyolipoma; RCC, renal cell carcinoma ACC, accuracy; AUC, the area under the receiver operating characteristic curve; Sen, sensitivity; Spe, specificity; Dice, Dice coefficient.