FCF-CSM: A Fuzzy Clustering Framework Based on Chromaticity Statistical Model for Automatic Segmentation of Port Wine Stains

Extracting lesions with accurate boundaries from clinical images is crucial for the clinical diagnosis, progression monitoring, and efficacy evaluation of port wine stains (PWS). However, accurately delineating lesion boundaries remains challenging due to complex boundary structures and unreliable annotations. In this paper, we propose a fuzzy clustering framework, FCF-CSM, for automated segmentation of PWS lesions. It combines prior knowledge of PWS color distribution with superpixels’ boundary characterization capability. Firstly, a chromaticity statistical model (CSM) is established based on 1000 collected PWS images, providing prior probabilities of PWS lesions to guide the improvement of the fuzzy clustering framework. Secondly, a superpixel method incorporating CSM is applied to PWS images, generating superpixels with improved boundary characterization. Thirdly, these superpixels are finely clustered using color features related to erythema index, color statistics, and color volume, improving PWS lesion distinction from complex backgrounds. Finally, a CSM-based automatic decision method distinguishes lesions from the background, achieving fully automated PWS segmentation within a fuzzy clustering framework. In addition, a boundary local fitting (BLF) metric is proposed to evaluate the segmentation precision of the PWS boundaries. Comparative experiments are conducted to verify the superiority of FCF-CSM. It achieves comparable overall segmentation performance with Jaccard and Dice metrics of 84.14% and 91.11%, respectively, compared to state-of-the-art methods. In terms of boundary segmentation, FCF-CSM outperforms other methods with an 81.52% BLF metric. FCF-CSM has proven to be effective for PWS segmentation and is promising to improve boundary delineation. The code is available at https://github.com/JinrongMu/FCF-CSM. Note to Practitioners—The motivation of this study was to construct a statistical model of port wine stain (PWS) color to quantify prior knowledge of PWS lesion color in RGB images. Existing methods for automatic segmentation of PWS rely heavily on annotated data, but the lack of publicly available datasets hinders the development of such algorithms due to the privacy of clinical data. This paper proposes a fuzzy clustering framework based on the chromaticity statistical model, which can achieve high-precision and fine-grained delineation of the boundaries of PWS lesions without annotating data. In this study, we describe the prior of PWS color distribution based on colorimetry theory and apply this knowledge to the PWS automatic segmentation task, thereby realizing knowledge sharing while protecting patient privacy from being leaked. Comprehensive comparative experiments demonstrate the effectiveness and reliability of the chromaticity statistical model. However, the dataset used to build this model only includes populations with yellow skin tones. In future research, we will address the automatic identification of PWS lesions applicable to other skin tones.