805 Segmentation of cutaneous chronic graft-versus-host disease by a deep learning neural network

Surface area involvement is a key measure for staging and tracking cutaneous chronic graft-versus-host disease (cGVHD) but limited by availability of expert dermatologist visual evaluation. We developed deep learning algorithm to segment (mark out) skin affected cGVHD tested its performance on unseen patients. took 179 181 unaffected 3D photos from 36 As ground truth, human annotator marked each region cGVHD. A U-Net was trained in leave-one-patient-out validation experiment. Without knowing board-certified assessed segmentation clinical relevance. The Dice coefficient used quantify spatial overlap. error quantitatively estimate error. In the patient, identified with overall accuracy, positive predictive value, negative values exceeding 90%. 77% segmentations were rated clinically acceptable excellent. At pixel-level, relative achieved median 0.74 (interquartile range: 0.40 – 0.89), surface 8.89% (22.10 3.69%). significant difference observed between Fitzpatrick types I-III compared IV-VI (median 7.17% vs 24.06% respectively). No found patients erythematous or sclerotic disease. provided majority this study. Automated algorithms could provide consistent method score track involvement.