Discriminative deep learning based benignity/malignancy diagnosis of dermatologic ultrasound skin lesions with pretrained artificial intelligence architecture

• DOI: 10.1111/srt.13086
• Publication Date: 2021-08-22T11:58:48Z
• Authors: Alexandra Laverde...; Abdulhadi Jfri; Rubén García; Irene Salgüero; Constanza Martínez; Hirune Cembrero; Gastón Roustán; Fernando Alfageme
• Keywords: Skin Neoplasms; dermatologic ultrasound; deep learning; Original Articles; artificial intelligence; Sensitivity and Specificity; 3. Good health; 03 medical and health sciences; Deep Learning; 0302 clinical medicine; Artificial Intelligence; skin ultrasound; Humans; Neural Networks, Computer; Ultrasonography

AbstractBackgroundDeep‐learning algorithms (DLAs) have been used in artificial intelligence aided ultrasonography diagnosis of thyroid and breast lesions. However, its use has not been described in the case of dermatologic ultrasound lesions. Our purpose was to train a DLA to discriminate benign form malignant lesions in dermatologic ultrasound images.Materials and methodsWe trained a prebuilt neural network architecture (EfficientNet B4) in a commercial artificial intelligence platform (Peltarion, Stockholm, Sweden) with 235 color Doppler images of both benign and malignant ultrasound images of 235 excised and histologically confirmed skin lesions (84.3% training, 15.7% validation). An additional 35 test images were used for testing the algorithm discrimination for correct benign/malignant diagnosis. One dermatologist with more than 5 years of experience in dermatologic ultrasound blindly evaluated the same 35 test images for malignancy or benignity.ResultsEfficientNet B4 trained dermatologic ultrasound algorithm sensitivity; specificity; predictive positive values, and predicted negative values for validation algorithm were 0.8, 0.86, 0.86, and 0.8, respectively for malignancy diagnosis. When tested with 35 previously unevaluated images sets, the algorithm´s accuracy for correct benign/malignant diagnosis was 77.1%, not statistically significantly different from the dermatologist's evaluation (74.1%).ConclusionAn adequately trained algorithm, even with a limited number of images, is at least as accurate as a dermatologic‐ultrasound experienced dermatologist in the evaluation of benignity/malignancy of ultrasound skin tumor images devoid of clinical data.