Development and validation of the interpretability analysis system based on deep learning model for smart image follow-up of nail pigmentation.

BACKGROUND

Nail pigmentation can be a clinical manifestation of malignant melanoma and a variety of benign diseases. Nail matrix biopsy for pathologic examination, the gold standard for diagnosis of subungual melanoma, is a painful procedure and may result in nail damage. Therefore, there is a great need for non-invasive methods and long-term follow-up for nail pigmentation. The objective of this study is to establish an intelligent precursor system to provide convenient monitoring for nail pigmentation, early warning subungual melanoma, and reduce nail biopsies to the minimum necessary.

METHODS

Dermoscopic images of nail lesions were obtained from outpatients between 2019 and 2020. The images were divided into the training set and the test set using k-fold cross validation at a ratio of 10:1. The deep learning model is developed based on the Pytorch framework. The model structure is optimized using the image segmentation model U-Net. An image segmentation module analyzed the contours of the whole nail plate and pigmented area according to the boundary features of the input images and a rule calculation module used the output information of the segmentation model to automatically analyze specific indicators referring to the "ABCDEF" rule. The model's results were compared with those of clinical experts.

RESULTS

From 550 dermoscopic images of nail lesions obtained, 500 were selected randomly as the training set, and the remaining 50 as the test set. Our image segmentation module realized automatic segmentation of the pigmented area and the whole nail plate with dice coefficient to be 0.8711 and 0.9652, respectively. Five qualitative indicators were selected in the interpretability analysis system and the models showed a certain degree of consistency with the evaluation by clinical experts, i.e., R2 for area ratio vs. breadth score was 0.8179 (P<0.001), for mean pixel value vs. pigment score was 0.7149 (P<0.001), for evenness vs. pigment score was 0.5247 (P<0.001).

CONCLUSIONS

The proposed system made accurate segmentation of the nail plate and pigmented area and achieved medically interpretable index analysis. It is potentially a primer of an intelligent follow-up system that will enable convenient and spatially unaffected management and monitoring of nail pigmentation. It may provide clinicians with understandable auxiliary information for diagnosis.