Source-linked AI summary
Lesion Border Detection in Dermoscopy Images Using Ensembles of Thresholding Methods
M. Emre Celebi, Quan Wen, Sae Hwang, Hitoshi Iyatomi, Gerald Schaefer
TL;DR
Difficult and subjective visual interpretation motivates automated dermoscopy analysis, with border detection serving as an important early step. The paper fuses several thresholding methods for lesion-border detection and reports favorable performance on 90 difficult images compared with nine recent methods, while noting vulnerability to hair and bubbles.
Problem
Visual interpretation of dermoscopy images is difficult and subjective, while no single thresholding method robustly handles varied image characteristics.
Method
The method fuses several thresholding methods applied to the blue channel, followed by simple postprocessing to obtain lesion borders.
Results
On 90 dermoscopy images, the method compared favorably with nine recent border-detection methods and required 0.1 seconds for a typical 768 × 512 image.
Takeaways & Limitations
Threshold fusion provides a fast, accurate, and stable border-detection approach that is comparable to the best thresholding method across image characteristics.
Takeaways & Limitations
The method may perform poorly on images with substantial hair or bubbles because these elements alter histograms and bias threshold computations.
Abstract
from arXiv · showhide
Dermoscopy is one of the major imaging modalities used in the diagnosis of melanoma and other pigmented skin lesions. Due to the difficulty and subjectivity of human interpretation, automated analysis of dermoscopy images has become an important research area. Border detection is often the first step in this analysis. In many cases, the lesion can be roughly separated from the background skin using a thresholding method applied to the blue channel. However, no single thresholding method appears to be robust enough to successfully handle the wide variety of dermoscopy images encountered in clinical practice. In this paper, we present an automated method for detecting lesion borders in dermoscopy images using ensembles of thresholding methods. Experiments on a difficult set of 90 images demonstrate that the proposed method is robust, fast, and accurate when compared to nine state-of-the-art methods.
1 Introduction
Melanoma diagnosis benefits from dermoscopy, but visual interpretation remains difficult and subjective, motivating computerized analysis. Border detection is an important early step because lesion borders support diagnosis and extraction of clinical features.
- Early melanoma detection matters because the disease can be cured with simple excision when identified early.
- Dermoscopy uses magnified, specialized lighting to reveal subsurface structures and morphological features more clearly than conventional clinical images.
- Computerized image analysis is important because visual interpretation of dermoscopy images is difficult and subjective.
- Border detection is an early analysis step whose accuracy supports diagnosis and extraction of features such as pigment networks, globules, and blue-white areas.
- Border detection is challenging because lesions may have low-contrast, fuzzy, irregular borders and artifacts, variegated coloring, or fragmentation.
- Existing approaches span thresholding, clustering, deformable models, segmentation, dynamic programming, and supervised learning; this paper proposes fusing thresholding methods with simple postprocessing.
2 Threshold Fusion
The method fuses multiple thresholding results because individual methods vary with image statistics, then applies simple postprocessing to obtain lesion borders. The fusion is formulated within an MRF framework and uses threshold outputs, energy minimization, and morphological expansion.
- Motivation: Thresholding effectiveness depends strongly on image statistics, so a method that performs well generally may fail on a particular dermoscopy image.In the illustrated case, method 1(e) performs well, while methods 1(c), 1(d), and 1(f) underestimate the optimal threshold; method 1(f) performs worst despite its popularity.
- Motivation: An ensemble fuses the outputs of several thresholding methods to seek robust decisions across varying image characteristics.The stated goal is to achieve accuracy comparable to the best individual method without requiring the same method to work best for every image.
- MRF formulation: The fusion treats the binary output as an MRF energy-minimization problem using spatial and inter-image energy terms.The spatial term uses a predefined neighborhood system, while inter-image terms weight the participating thresholding methods.
- Fusion procedure: Each ensemble method is applied to the image to generate thresholded images, after which the fusion labels are initialized without and then updated with spatial energy.The full iterative spatial-update procedure can be repeated up to Kmax times, but this study uses only the first two steps because later iterations made only a marginal contribution.
- Postprocessing: The final border is formed by filling the binary fusion output and retaining only its largest 4-connected component.Morphological dilation then expands the fusion output using a circular structuring element whose radius depends on the lesion diameter; in the reported example, XOR error falls from 17.14% to 8.79%.
3 Experimental Results and Discussion
The proposed fusion method was evaluated on 90 dermoscopy images against nine recent methods, showing lower and more stable XOR errors. Individual thresholding methods had higher mean errors and greater vulnerability to pathological images.
- Evaluation: 90 dermoscopy images, including 23 invasive malignant melanoma and 67 benign lesions, were evaluated against nine recent border-detection methods using percent XOR error.Manual borders were drawn by an experienced dermatologist; XOR error measures disagreement between automated and manual binary outputs.
- Overall comparison: The fusion method was significantly more accurate than the other automated methods and had low standard deviation, indicating greater stability.Table 1 reports mean and standard deviation percent XOR errors for ten automated methods.
- Individual methods: Individual thresholding methods produced significantly higher mean errors than fusion because pathological images made them more prone to catastrophic failures.Their high standard deviations also support this explanation.
- Interpretation: The fusion goal was to achieve performance comparable to the best thresholding method independently of image characteristics, rather than outperform every individual method.
- Qualitative results: Sample results show good performance despite diffuse edges, blood vessels, and skin lines.
4 Conclusions
The paper concludes that threshold fusion provides favorable lesion-border detection compared with nine recent methods while remaining easy to implement and fast. Its performance may degrade on images containing substantial hair or bubbles because these alter histogram-based threshold computations.
- Conclusion: The automated threshold fusion method compares favorably with nine recent border-detection methods on a difficult image set.
- Conclusion: 0.1 seconds is the reported processing time for a typical 768 × 512 image on an Intel QX9300 2.53GHz computer.The method is also described as easy to implement and requiring no sophisticated postprocessing.
- Limitations: Images with significant hair or bubbles may reduce performance because these elements alter the histogram and bias threshold computations.A reliable bubble-removal method remains an open problem.