Tre-O1-04 - Assessment of skin involvement in cutaneous T-cell lymphoma using convolutional neural networks

Abstract

<h3>Introduction:</h3> Mycosis fungoides (MF) and Sézary syndrome (SS) are the most common types of cutaneous T-cell lymphomas (CTCL). The assessment of skin involvement is critical for the correct TNMB classification, however, prone to intra- and interobserver variability. The aim of this project was an automated identification of skin lesions with artificial intelligence on clinical images of CTCL patients to reduce subjectivity in the evaluation of skin lesions. <h3>Materials and methods:</h3> The analysis included two datasets. The first set included 50 standard photographic clinical images of three CTCL patients with multilocular disease. These images were manually segmented by a dermatologist to obtain binary masks used as ground truth. The second set included 50 lesional images with a calibration marker of 15 patients retrieved with mobile devices (iPhone, iPad) and manually marked for lesional areas on the device. For both cohorts, all images were split with a ratio of 80:20 into training and validation set and analyzed using a convolutional neural network (CNN) created in cooperation by the imito AG, Zürich, the department of informatics of the University Hospital and Institute of Medical Informatics at Luebeck University. <h3>Results:</h3> Overall, 15 CTCL patients (11 M, 4 F; M:F 2.75), median age seventy (41-80 years) were included. 13 patients were diagnosed with MF, stages IA (3), IB (5), IIB (4), IIIA (1) and two patients with SS. The median mSWAT was 8 (ranges 1–85). Due to the large image size, all images were split into smaller pixel sizes, so that the final image set included 931 images for the standard photographs and 1123 images in the mobile device cohort. The CNN was created as a UNet and trained with a Google OpenImages dataset. Several analyses were performed and the finally used variant chosen based on the test accuracy (F-score and Jaccard index). Within the first set of images, lesional areas were correctly identified in 79.8% and non-lesional in 89.2%, compared to 19.8% and 96.8% in the second set. The lesional areas were digitally calculated in the second set. <h3>Conclusion:</h3> The image classification in the first set was considered very convincing. However, the discrepancy between the results of the analysis of the first and second set appear striking with a much better classification and correct identification of lesional areas in the first set. Both approaches were based on trained algorithms as explained. Therefore, we presume that the inferior recognition in the mobile cohort is most likely due to a less exact mark-up of the lesions on the device and the restriction that only one lesion could be marked as a lesion. Overall, we consider that the standardized assessment of cutaneous involvement in CTCL may be helpful in the evaluation of skin lesions in daily practice and in clinical trials. With further development, an automated full-body skin lesion recognition and calculation of involved areas might be feasible.