Can convolutional neural networks outperform clinicians in the detection of melanoma on dermoscopy images?

Introduction: Artificial intelligence is widely used in various _elds of medicine. It also has great potential for being used in the assessment of dermoscopy images. This study aimed to evaluate whether a convolutional neural network model could match dermatologists’ accuracy in the assessment of dermoscopic pictures. Material and methods: For this research we used HAM10000 training dataset, that was extracted from “ISIC 2018: Skin Lesion Analysis Towards Melanoma Detection”. All skin lesions were classi_ed to one of the following group: (1) malignant melanoma, (2) melanocytic nevus, (3) basal cell carcinoma, (4) actinic keratosis/Bowen’s disease, (5) benign keratosis, (6) dermato_broma, and (7) vascular lesion. From the dataset, we have randomly extracted 104 images from all classes of lesions to create the online test presented to 14 dermatologists who were asked to classify each lesion out of 104 dermoscopic pictures to the groups mentioned above. Next, the ResNeXt model was evaluated on the same dataset. Results: Dermatologists achieved better sensitivity than ResNeXt in malignant melanoma di_erentiation. However, precision and F1 score of ResNeXt were higher in comparison to dermatologists. Moreover, CNN was more precise and sensitive to other skin lesion types. Conclusions: This research has shown that computer vision aided dermoscopy can be a supportive tool that could help physicians in the screening of patients for malignant melanoma. Introduction: Artificial intelligence is widely used in various _elds of medicine. It also has great potential for being used in the assessment of dermoscopy images. This study aimed to evaluate whether a convolutional neural network model could match dermatologists’ accuracy in the assessment of dermoscopic pictures. Material and methods: For this research we used HAM10000 training dataset, that was extracted from “ISIC 2018: Skin Lesion Analysis Towards Melanoma Detection”. All skin lesions were classi_ed to one of the following group: (1) malignant melanoma, (2) melanocytic nevus, (3) basal cell carcinoma, (4) actinic keratosis/Bowen’s disease, (5) benign keratosis, (6) dermato_broma, and (7) vascular lesion. From the dataset, we have randomly extracted 104 images from all classes of lesions to create the online test presented to 14 dermatologists who were asked to classify each lesion out of 104 dermoscopic pictures to the groups mentioned above. Next, the ResNeXt model was evaluated on the same dataset. Results: Dermatologists achieved better sensitivity than ResNeXt in malignant melanoma di_erentiation. However, precision and F1 score of ResNeXt were higher in comparison to dermatologists. Moreover, CNN was more precise and sensitive to other skin lesion types. Conclusions: This research has shown that computer vision aided dermoscopy can be a supportive tool that could help physicians in the screening of patients for malignant melanoma.