Table 2.
The application of AI in multi-classification for skin lesions.
| Authors | Refer ence | Year | Country | AI Algorithm Model | The Purpose of AI Algorithm | Image (Datasets) Recourse | No. of Images in Datasets | Usage | Types of Images | Accuracy /Precision (%) |
Sensitivity/Recall (%) | Specificity (%) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Kassem et al. | [23] | 2020 | Egypt | Deep CNNs (modified GoogleNet) | Classification of multiple skin lesions | ISIC 2016–2019 | 25,331 | Multi-class (8) | Dermoscopy | 94.92 | 79.8 | 97 |
| Rezvantalab et al. | [52] | 2018 | Iran | Four deep learning convolutional neural networks (CNNs) | Investigating the ability of deep convolutional neural networks in classification of multiple skin lesion | HAM10000; PH2 | 10,135 | Multi-class (8) | Dermoscopy | 80.22–89.01 | 82.26–99.10 | 79.60–89.01 |
| Gessert et al. | [53] | 2018 | German y | Ensemble of CNN | Diagnosis of multiple skin lesions | ISIC-2018, HAM10000 | 23,515 | Multi-class (7) | Dermoscopy | 85.1 | 93.1–97.6 | N/A |
| Gessert et al. | [54] | 2020 | German y | Ensemble of multi-resolution CNN | Classification of multiple skin lesions | HAM10000, BCN20000, MSK,7- point, |
47,049 | Multi-class (8) | Dermoscopy | 80.5–96 | 72.5–74.2 | 94–99.9 |
| Haenssle et al. | [55] | 2018 | German y | Deep convolutional neural network (Google’s Inception v4 architecture) | Detection of melanoma and comparison of its performance with 58 dermatologists | ISIC archive, clinical images | >150,000 | Multi-class (20) | Macroscopy and Dermoscopy |
86 | 86.6–88.9 | 71.3–75.7 |
| Haenssle et al. | [56] | 2020 | Multi- country | FotoFinder® Moleanalyzer Pro | Classification of skin lesions and comparison of the performance of the AI model with 96 dermatologists | ISIC archive, clinical images | >150,000 | Multi-class (25) | Macroscopy and Dermoscopy |
84 | 95 | 76.7 |
| Esteva et al. | [66] | 2017 | USA | Deep convolutional neural networks (GoogleNet Inception v3) | Classification of skin cancer and comparison of the performance of AI model with 21 dermatologists | Online repositories and clinical data from | 129,450 | Multi-class (2032) | Macroscopy and Dermoscopy |
1: 72.1 ± 0.9; 2: 55.4 ± 1.7 |
N/A | N/A |
| Mahbod et al. | [67] | 2020 | Austria | Multi-scale multi-convolutional neural networks (MSM-CNNs) | Investigating the effect of image size for skin lesion classification | ISIC-2016, 2017, 2018 HAM10000 | 12,927 | Multi-class (7) | Dermoscopy | 96.3 | N/A | N/A |
| Iqbal et al. | [71] | 2020 | China | Deep CNN | Classification of multiple skin lesion | ISIC-2017, 2018, 2019 | 25,331 | Multi-class (8) | Dermoscopy | 94 | 93 | 91 |
| Qin et al. | [73] | 2020 | China | Generative adversarial networks (GANs) | Classification of multiple skin lesion | ISIC-2018 | 10,015 | Multi-class (7) | Dermoscopy | 95.2 | 83.2 | 74.3 |
| Cano et al. | [74] | 2021 | Panama | NasNet | Classification of multiple skin lesions | ISIC-2019 | 25,331 | Multi-class (8) | Dermoscopy | 71–99 | 73–98 | 70–99 |
| Barhoumi et al. | [75] | 2021 | Tunisia | Transfer learning CNN model | Classification of multiple skin lesions | ISIC 2018 | 5057 | Multi-class (7) | Dermoscopy | 95 | 96 | N/A |
| Ratul et al. | [76] | 2020 | Canada | Dilated CNNs (VGG-16,-19, MobileNet, Inception-V3) | Classification of multiple skin lesions | HAM10000 | 10,015 | Multi-class (7) | Dermoscopy | 87–89 | 87–89 | N/A |
| Rashid et al. | [77] | 2020 | Pakistan | Semi-supervised GANs | Classification of multiple skin lesions | ISIC 2018 | 10,000 | Multi-class (7) | Dermoscopy | 73–94 | 69–92 | N/A |
| Maron et al. | [78] | 2019 | German y | CNNs | Classification of multiple skin lesions and comparison of the performance of the AI model with 112 dermatologists | ISIC 2018, HAM10000 | 11,444 | Multi-class (5) | Dermoscopy | N/A | 90.2–97.7 | 94.2–99.5 |
| Sun et al. | [79] | 2021 | China | CNNs | Classification of multiple skin lesions | ISIC-2019, MED- NODE, PH2, 7- point |
18,460 | Multi-class (7) | Dermoscopy | 66.2–89.5 | 66.2–89.5 | 95.2–99.3 |
| Jain et al. | [80] | 2021 | India | Six transfer learning nets | Classification of multiple skin lesions | HAM10000 | 10,015 | Multi-class (7) | Dermoscopy | 66–90 | 66–90 | N/A |
| Winkler et al. | [81] | 2020 | Gemany | FotoFinder® Moleanalyzer Pro (CNN) | Detection of various melanoma localizations and subtypes | ISIC archive, clinical images | >150,000 | Multi-class (6) | Macroscopy and Dermoscopy |
50.8–95.4 | 53.3–100 | 65–94 |
| Binder et al. | [82] | 1994 | Austria | Artificial neural networks (ANNs) | Classification of naevi and malignant melanoma and comparison of the performance of AI model with 3 dermatologists | Oil immersion images of pigmented skin lesions | 200 | Multi-class (3) | Microscopy | 86 | 95 | 88 |
| Sies et al. | [83] | 2020 | German y | FotoFinder® Moleanalyzer Pro/FotoFinder®Moleanalyzer- 3, Dynamole | Detection of various melanoma localizations and subtypes | ISIC dermoscopic archive, multicentric clinical images | >150,000 | Multi-class (20) | Dermoscopy | 92.8 | 77.6 | 95.3 |
| Yang et al. | [84] | 2020 | China | CNNs (DenseNet-96, ResNet-152, ResNet-99) | Classification of multiple benign hyperpigmented dermatitis and comparison of the performance of AI model with 11 dermatologists | Clinical images | 12,816 | Multi-class (6) | Macroscopy | 75.3–97.8 | 75.5–94.4 | 95.6–99.8 |
| Lyakhov et al. | [85] | 2022 | Russia | Multimodal neural network | Recognition of multiple pigmented skin lesions | ISIC-2016–2021 | 41,725 | Multi-class (10) | Dermoscopy | 83.6 | N/A | N/A |
| Guzman et al. | [86] | 2015 | Philippin es | Singe/multi-level and multi-models ANN | Detection of eczema skin lesion | Clinical images | 504 | Multi-class (3) | Macroscopy | Single: 78.17–87.30 Multi: 81.34–85.71 |
N/A | N/A |
| Han et al. | [87] | 2018 | Korea | Region-based convolutional deep neural networks | Diagnosis of onychomycosis and comparison of the performance of AI model with 42 dermatologists | Clinical images | 49,567 | Multi-class (6) | Macroscopy | 82–98 | 82.7–96 | 69.3–96.7 |
| A.Blum et al. | [88] | 2004 | Gemany | Vision algebra algorithms | Diagnosis of melanocytic lesions and validation of its diagnostic accuracy | Clinical images | 837 | Multi-class (20) | Dermoscopy | 82.3–84.1 | 80–88.1 | 82.4–82.7 |
| Marchetti et al. | [89] | 2020 | USA | CNNs and deep learning algorithms | Classification of melanoma and comparison of the performance of AI model with 17 dermatologists | ISIC-2017 | 2750 | Multi-class (3) | Dermoscopy | 86.8 | 76 | 85 |
| Shen et al. | [90] | 2018 | China | Convolutional neural networks | Diagnosing for facial acne vulgaris | Clinical images | Binary: 6000 Multi:42,000 |
Binary-class/Multi-class (7) | Macroscopy | 88.7–89.5 | 81.7–92 | 87–95.7 |
| Seité et al. | [91] | 2019 | France | Deep learning algorithm | Determination of the severity of facial acne and identification of subtypes of acne lesion | Clinical images | 4958 | Multi-class (3) | Macroscopy | N/A | N/A | N/A |
| Zhao et al. | [92] | 2019 | China | CNNs | Identification of psoriasis | XiangyaDerm-Pso9 | 8021 | Multi-class (9) | Macroscopy | 88 | 83–95 | 96–98 |
| Han et al. | [93] | 2020 | Korea | Deep Neural Networks | Predicting malignancy and suggesting treatment option, as well as multi-classification for 134 skin disorders | Clinical images | 220,680 | 1:Binary- class 2:Multi-class (134) |
Macroscopy | 1: 56.7–92 2: 44.8–78.1 |
N/A | N/A |