[Paper Review] Automatic Skin Lesion Segmentation using Semi-supervised Learning Technique

S. M. Jaisakthi, Chandrabose Aravindan+1 more|arXiv (Cornell University)|Mar 13, 2017

Cutaneous Melanoma Detection and ManagementMedicine1 references22 citations

TL;DR

This paper proposes a semi-supervised skin lesion segmentation method using preprocessing for illumination correction and hair removal, followed by k-means clustering and histogram-based color filtering to isolate lesions. The approach achieved a Jaccard score of 0.548 on the ISIC 2017 validation set, demonstrating effectiveness in automatic lesion detection for early melanoma diagnosis.

ABSTRACT

Skin cancer is the most common of all cancers and each year million cases of skin cancer are treated. Treating and curing skin cancer is easy, if it is diagnosed and treated at an early stage. In this work we propose an automatic technique for skin lesion segmentation in dermoscopic images which helps in classifying the skin cancer types. The proposed method comprises of two major phases (1) preprocessing and (2) segmentation using semi-supervised learning algorithm. In the preprocessing phase noise are removed using filtering technique and in the segmentation phase skin lesions are segmented based on clustering technique. K-means clustering algorithm is used to cluster the preprocessed images and skin lesions are filtered from these clusters based on the color feature. Color of the skin lesions are learned from the training images using histograms calculations in RGB color space. The training images were downloaded from the ISIC 2017 challenge website and the experimental results were evaluated using validation and test sets.

Motivation & Objective

• To develop an automatic skin lesion segmentation system that improves early detection of melanoma.
• To address challenges in dermoscopic image analysis such as uneven illumination, hair artifacts, and noise.
• To leverage semi-supervised learning by using ground truth from training data to guide lesion segmentation.
• To enhance clinical accuracy by isolating lesion regions using color and clustering-based segmentation.

Proposed method

• Images were resized to 25% of original size if dimensions exceeded 1500 pixels to reduce computational load.
• Illumination correction was performed using CLAHE on the L channel of Lab color space to normalize lighting variations.
• Hair removal was achieved using the Frangi vesselness filter, which enhances tubular structures based on Hessian matrix eigenvalue analysis.
• Color features were extracted using RGB histograms from ground truth training images to define lesion color boundaries.
• K-means clustering with k=5 was applied to group pixels by color similarity, separating lesion regions from background and artifacts.
• Flood fill with two seed points (center and near boundary) was used to segment dense and boundary lesion regions, followed by connected component labeling.

Experimental results

Research questions

• RQ1 Can semi-supervised learning improve skin lesion segmentation accuracy using limited annotated training data?
• RQ2 How effective is k-means clustering combined with color histogram analysis in distinguishing skin lesions from normal tissue and artifacts?
• RQ3 To what extent do preprocessing steps like CLAHE and Frangi filtering enhance segmentation performance?
• RQ4 What is the impact of using multiple seed points in flood fill on lesion region completeness?

Key findings

• The proposed method achieved a Jaccard score of 0.548 on the ISIC 2017 validation set, indicating moderate to strong lesion overlap with ground truth.
• Sensitivity was high, suggesting effective detection of actual lesion regions despite low Jaccard index.
• The use of CLAHE significantly improved contrast and uniformity in lesion regions by correcting non-uniform illumination.
• Frangi vesselness filtering effectively reduced false positives by removing hair and other tubular artifacts.
• K-means clustering with k=5 successfully separated lesion regions from background, noise, and imaging artifacts.
• The dual-seed flood fill approach improved segmentation completeness by capturing both central and boundary lesion areas.