Table 1. Summary of descriptive characteristics of included articles (n = 09).
| Author, year, country and design studies |
Included Studies | Type of cancer | Index test | Reference test | True positives / N of images | True Negatives /N of images | Sensitivity and Specificity/ odds ratio, Mean±SD, p’ value |
Diagnostic accuracy (%), Mean±SD, p’ value | Conclusions |
|---|---|---|---|---|---|---|---|---|---|
| Dorrius et al, 2011 [39], Netherlands, Descriptives studies |
10 | Breast Cancer | Computer-aided-detection (CADe) | Magnetic Resonance Imaging (MRI) | - | - | Sensitivity Radiologist no CAD, general 82% (95% CI: 72%–90%) Radiologist with CAD, general 89% (95% CI: 83%–93%) Specificity Radiologist no CAD, general 81% (95% CI: 74%–87%) Radiologist with CAD, general 81% (95% CI: 76%–85%) |
- | MR images CAD has little influence on the sensitivity and specificity of the performance of radiologists experienced in breast MRI diagnosis. Breast MRI interpretation by radiologists remains essential. Radiologists with less experience seem to benefit from a CAD system when performing breast MRI evaluation. |
| Henriksen EL et al, (2018) [37], Denmark Clinical trials |
13 | Breast cancer | CAD system.; Single Reading (SR) SR vs SR + CAD; Double Reading (DR) DR vs SR þ CAD; |
MM | - | - | - | - | In conclusion, all but two studies found that SR CAD improves mammography screening RRs, sensitivity, and CDR when compared to SR alone. No statistically significant variations in sensitivity or CDR were seen when compared to DR. More research is needed to assess the impact of CAD in a population-based screening program with high-volume readers. Longer follow-up studies are required for a thorough assessment of cancer rates. And studies based on digital mammography are required to assess the efficacy of CAD in the current standard of care technology. |
| Nindrea et al, 2018 [43], Indonesia, Diagnostic Accuracy studies |
11 | Breast cancer | Machine Learning Algorithms Super Vector Machine (SVM); Artificial Neural Networks (ANN); Decision Tree (DT); Naive Bayes (NB); K-Nearest Neighbor (KNN) |
Mammography (MM) | SVM 40,37%/3532; ANN 1,30%/63325 DT 33,19%/738 NB 35,32%/1039 KNN 41%/1568 |
SVM 46,40%/3532 ANN 97,88%/63325 DT 61,38%/738 NB 54,66%/1039 KNN 44,89%/1568 |
Sensitivity SVM: 0.67–0.99 (95% CI: ([0.41–0.87]-[0.95–1.00]); ANN: 0.84–0.97 (95% CI: ([0.60–0.97]-[0.95–98]); DT: 0.90–0.92 (95% CI: ([0.68–0.99]-[0.88–.95]); NB: 0.76–0.91 (95% CI: ([0.68–0.83]-[0.87–.95]); KNN: 0.56–0.95 (95% CI: ([0.48–0.64]-[0.92–0.97]); Specificity SVM: 0.60–0.98 (95% CI: ([0.36–0.81]-[0.96–1.00]); ANN: 0.71–0.99 (95% CI: ([0.48–0.89]-[0.99–0.99]); DT: 0.79–0.97 (95% CI: ([0.54–0.94]-[0.9–0.98]); NB: 0.78–0.99 (95% CI: ([0.52–0.94]-[0.9–1.00]); KNN: 0.53–0.99 (95% CI: ([0.44–0.61]-[0.93–0.97); |
SVM: 99.51%; ANN: 97.3%; DT: 95.13%; NB: 95.99%; KNM: 95.27%; |
Therefore, the early diagnosis of breast cancer will be more effective, and the mortality rate of breast cancer will decrease. Additionally, if the present method is designed in the form of a web-based or smartphone application, women who want to know their own risk of breast cancer will be able to access this information easily in daily life. |
| Azavedo et al, 2012 [35], Sweden, Prospective or Retrospective studies |
4 | Breast cancer | Computer-aided-detection (CAD) | MM | - | - | - | - | The scientific evidence is insufficient to determine whether CAD + single reading by one breast radiologist would yield results that are at least equivalent to those obtained in standard practice, i.e. double reading where two breast radiologists independently read the mammographic images. |
| Eadie et al, 2012 [44], United Kingdom, Diagnostic Accuracy studies |
48 | Breast cancer, lung cancer, liver cancer, prostate cancer, bone cancer, bowel cancer, skin cancer, neck cancer. |
CADe; Diagnostic CAD (CADx) |
MM; Breast ultrasound (BUS); BUS + mammogram; Lung Conputered Tomography (LCT); Dermatologic; |
- | - | Sensitivity (SD) CADe overall Radiologist alone: 80.41±1.46 With CAD: 84.02±1.30 CADx overall Radiologist alone: 2.79±6.12 With CAD: 90.66±4.07 Specificity (SD) CADe overall Radiologist alone: 90.10±1.97 With CAD: 87.08±2.75 CADx overall Radiologist alone: 83.00±14.46) With CAD: 88.04±15.03 |
Diagnostic odds ratio (DOR) (SD) CADe overallRadiologist alone3.63±0.16With CAD:3.58±0.20CADx overallRadiologist alone3.44±0.79With CAD: 4.75±0.91 |
Certain types of CAD did offer diagnostic benefit compared with radiologists diagnosing alone: significantly better ln DOR scores were seen with CADx systems used with mammography and breast ultrasound. Applications such as lung CT and dermatologic imaging do not seem to benefit overall from the addition of CAD. These findings therefore offer suggestions about how CAD can be best applied in the diagnosis of cancer using imaging. |
| Zhao et al, 2019 [42], China, Prospective or Retrospective studies |
5 | Thyroid (nodules) cancer | CADx system |
US | positive likelihood ratio CADx system 4.1 (95% CI 2.5–6.9); CADx by Samsung 4.9 (95% CI 3.4–7.0); radiologists 11.1 (95% CI 5.6–21.9); |
negative likelihood ratio CADx sistem 0.17 (95% CI 0.09–0.32); CADx by Samsung 0.22 (95% CI 0.12–0.38); radiologists 0.13 (95% CI 0.08–0.21); |
Sensitivity CADx system 0.87 (95% CI: 0.73–0.94; I2 = 93.53%); CADx by Samsung 0.82 (95% CI: 0.69–0.91; I2 = 79.62%); radiologists 0.88 (95% CI: 0.80–0.93; I2 = 81.66%); Specificity CADx system 0.79 (95% CI: 0.63–0.89; I2 = 89.67%); CADx by Samsung 0.83 (95% CI: 0.76–0.89; I2 = 27.52%); radiologists 0.92 (95% CI: 0.84–0.96; I2 = 84.25%); |
DOR CADx system25 (95% CI: 15–42; I2 = 15.5%, p = 0.315);CADx by Samsung23 (95% CI: 11–46; I2 = 35.9%, p = 0 .197);radiologists86 (95% CI: 47–158; I2 = 41.1%, p = 0.147) |
The sensitivity of the CAD system in thyroid nodules was similar to that of experienced radiologists. However, the CAD system had lower specificity and DOR than the experienced radiologist. The CAD system may play the potential role as a decision-making assistant alongside radiologists in the thyroid nodules’ diagnosis. |
| Cuocolo et al, 2020 [40], Italy, Diagnostic Accuracy studies | 12 | PCa | Machine learning (ML) ANN; SVM; LDA; NB; Linear regression (LIR); Random forest (RF); Logistic regression (LOR); Convolutional neural network (CNN); Deep transfer learning (DTL); |
MRI | - | - |
ML in PCa identification–overall (95%CI: 0.81–0.91; I2 = 92%, p <0.0001); Biopsy group (95%CI: 0.79–0.91; I2 = 87%, p <0.0001); Radical prostatectomy group (95%CI: 0.76–0.99; I2 = 93%, p <0.0001); Deep learning (95%CI: 0.69–0.86; I2 = 86%, p = 0.0001); Non-deep learning (95%CI: 0.85–0.94; I2 = 89%, p <0.0001); |
AUC overall AUC = 0.86Biopsy groupAUC = 0.85;Rradical prostatectomy groupAUC = 0.88;Deep learningAUC = 0.78;Non-deep learningAUC = 0.90; |
The findings show promising results for quantitative ML-based identification of csPCa. The results suggest that the overall accuracy of ML approached might be comparable with that reported for traditional Prostate Imaging Reporting and Data System scoring. Nevertheless, these techniques have the potential to improve csPCa detection accuracy and reproducibility in clinical practice. |
| Tabatabaei et al, 2021 [36], USA Retrospectives studies |
18 | Glioma | DT; KNN; SVM; RF; LOR; LDA; LIR; Least Absolute Shrinkage and Selection Operator (LAS/SO); Elastic Net (EN); Gradient Descent Algorithm (GDA); Deep Neural Network (DNN) |
MRI | - | - | - | - | The results appear promising for grade prediction from MR images using the radiomics techniques. However, there is no agreement about the radiomics pipeline, the number of extracted features, MR sequences, and machine learning technique. Before the clinical implementation of glioma grading by radiomics, more standardized research is needed. |
| Xing et al, 2021 [41], China, Retrospective studies |
15 | prostate cancer (PCa); Peripheral zone (PZ); Transitional zone (TZ); Central gland (CG); |
CAD system.; ANN; SVM; Linear Discriminant Analysis (LDA); Radiomic Machine Learning (RML); Non—specific classifier (NSC); |
MRI | SVM 42,76%/608; ANN 34,55%/301; RML 34,78%/738; NSC 19,41%/1586; PZ 51,95%/256; TZ 59,67%/186; CG 32,39%/71; |
SVM 41,94%/608; ANN 37,54%/301; RML 32,60%/738; NPC 65,15%/1586; PZ 32,81%/256; TZ 26,34%/186; CG 46,47%/71; |
Sensitivity: 0.47 to 1.00 0.87(95% CI: 0.76–0.94; I2 = 90.3%, p = 0.00) ANN: 0.66 to 0.77 SVM: 0.87 to 0.92 LDA: NR RML: 0.96 Prostate zones PZ: 0.66 to 1.00 TZ: 0.89 to 1.00 CG: 0.66 Specificity: 0.47 to 0.89 0.76(95% CI: 0.62–0.85; I2 = 95.8%, p = 0.00) ANN: 0.64 to 0.92 SVM: 0.47 to 0.95 LDA: NR RML: 0.51 Prostate zones PZ: 0.48 to 0.89; TZ:0.38 to 0.85; CG:0.92 |
AUC 0.89 (95% CI: 0.86–0.91) |
The study indicated that the use of CAD systems to interpret the results of MRI had high sensitivity and specificity in diagnosing PCa. We believe that SVM should be recommended as the best classifier for the CAD system. |
Subtitles: CADe = Computer-aided-detection; MRI = Magnetic Resonance Imaging; SVM = Super Vector Machine; ANN = Artificial Neural Networks; DT = Decision Tree; NB = Naive Bayes; KNN = K-Nearest Neighbor; MM = Mammography; CADx = Diagnostic CAD; BUS = Breast ultrasound; DOR = Diagnostic odds ratio; LCT = Lung Conputered Tomography; CDR = CAD on cancer detection rate (CDR); DR = double reading; RR = Recall Rate; Pca = Prostate cancer; PZ = Peripheral zone; TZ = Transitional zone; CG = Central gland; LDA = Linear Discriminant Analysis; RML = Radiomic Machine Learning; NSC = Non—specific classifier; ML = Machine learningA; LIR = Linear regression; RF = Random forest; LOR = Logistic regression; CNN = Convolutional neural network; DTL = Deep transfer learning; LAS/SO = Least Absolute Shrinkage and Selection Operator; EN = Elastic Net; GDA = Gradient Descent Algorithm; DNN = Deep Neural Network; SR = Single Reading; DR = Double Reading.