Automated assessment of skin histological tissue structures by artificial intelligence in cutaneous melanoma

Abstract

Background

Prognostic histopathological features such as mitosis in melanoma are excluded from the staging systems due to inter-observer variability and time constraints. While digital pathology offers artificial intelligence-driven solutions, existing melanoma algorithms often underperform or narrowly focus on specific features, limiting their clinical utility.

Objective

Develop and validate an automated artificial intelligence-driven segmentation framework to identify multiple histological tissue structures within cutaneous melanoma images.

Methods

Employing 157 melanoma whole slide images, U-Net and DeepLab3+ classifiers were independently trained Oncotopix ® platform using manual annotations, to detect specific histological features, termed application. All the applications are progressively executable. The performance of each application was measured when both operating independently and with sequential detection when applied to ten independent validation set images using accuracy and F1-score as metrics. The model was further validated by applying it to 442 whole-slide melanoma images, with dermatopathologists reviewing the segmentation outputs.

Results

Seven applications were developed for progressive automated detection: Whole tissue (1) and tumour microenvironment (TME) (2), Hair follicles & sebaceous gland (3) within TME, ulceration (5), and melanoma cell area (6) based on DeepLab3+. Epidermis (4) and mitosis within the tumour area (7) based on U-Net. The applications demonstrated over 92 % accuracy and F1-score surpassing 80 %, except for the ulceration application (F1-score = 75 %). The pathologist examination indicated that 92 % of the 442 images had correct segmentations.

Discussion and Conclusion

The developed applications demonstrated high performance, enhancing the analysis of time-consuming histological features. The model facilitates the identification of histopathological features in large datasets allowing potential refinement of melanoma staging.