213 Novel transfer learning approach to achieve high prediction accuracy for skin cancer classification in imbalanced data sets

Non-invasive visual detection of skin cancers from benign tumors remains a challenge in clinical practice. Studies have claimed the non-inferiority artificial intelligence classifying common such as nevus and melanoma. Better algorithms are yet to be developed assist accurate diagnoses. The aim this current study is investigate whether small or limited sample size for AI training could achieve high diagnostic performance based on image-discrimination among tumors. Biopsy-proven images (melanoma, basal cell carcinoma, squamous seborrheic keratosis, nevus, other tumors, normal health controls (all unique images) were evenly distributed test sets establish tools 3 different convolutional neural network (CNN) deep learning modules (inception_v3, inception_resnet_v2, NASNET_large). Under task, CNN prediction sensitivity lesion at 88.4% 87.26% (inception_v3), 87.40% 86.49% (inception_rasnet_v2), 88.01% 87.64% (NASNET_large), respectively. F1-score each module 87.82%, 86.94%, 87.82% inception_v3, inception _rasnet_v2, NASNET_large, Our models demonstrated consistent exceptional accuracy all when used differentiate malignant lesions. With prudent principle inputting clear photos biopsy-proven diagnosis, automated AI-assisted highly exceeds standard dermatologist-level classification correct diagnosis very size. This computer-assisted technique with displays its practicality largely spare an invasive biopsy acquire more confirmed useful tool support clinician's evaluation.