Abstract
Skin lesion border irregularity, representing the "B" feature in the ABCD rule, is a critical factor in melanoma diagnosis. However, its assessment often relies on subjective clinical judgment. While Artificial Neural Networks (ANNs) have advanced supervised learning, they typically fail to account for input ambiguity during training, which can degrade performance. This paper proposes a Fuzzy Multilayer Perceptron (F-MLP) that incorporates input ambiguity to refine the learning process. We introduce a fuzzy gradient descent optimization function and a type-II fuzzy sigmoid activation function. The F-MLP was applied to predict border irregularity after automated segmentation and feature extraction. Results show that the F-MLP outperforms standard neural networks and several state-of-the-art classifiers, though it requires more training time.
Introduction
The diagnosis of melanoma frequently utilizes the ABCD rule: Asymmetry, Border irregularity, Color variegation, and Diameter. Among these, border irregularity is a primary indicator of malignancy. Despite its importance, clinical assessment is notoriously subjective, leading to significant inter-observer variability. Automated, objective methods are therefore essential for reliable diagnosis.
Standard multilayer perceptrons (MLPs) treat all training samples with equal certainty. In medical imaging, however, many samples are ambiguous—occupying a "gray area" between classes. When a standard MLP updates its weights based on these uncertain inputs, it may learn misleading patterns. To address this, we propose a fuzzy-logic-enhanced MLP (F-MLP) designed to handle the inherent uncertainty in dermoscopic images.
Methodology
Image Processing Pipeline
The proposed system follows a multi-stage approach to move from raw pixels to a diagnostic prediction:
- Segmentation: Lesions are first segmented from the skin using a U-Net architecture.
- Edge Detection: The lesion border is extracted using FuzzEdge, a fuzzy-logic-based edge detector previously developed by the authors.
- Feature Extraction: Two sophisticated descriptors quantify the border's irregularity:
- Fractal Dimension: Measures the complexity and "raggedness" of the border.
- Convexity: The ratio of the lesion's perimeter to its convex hull perimeter.
The F-MLP Architecture
The core innovation is the F-MLP, which integrates fuzzy logic into the neural network's learning and activation mechanisms to dampen the influence of ambiguous data.
Fuzzy Gradient Descent
We modified the standard gradient descent to include an "ambiguity parameter." Using Fuzzy C-Means, we calculate membership degrees ($u$) for each input. The weight update is weighted by the difference in membership degrees ($|u_1 - u_2|^2$). If a sample is highly ambiguous (e.g., $u_1 \approx u_2 \approx 0.5$), its influence on the weight update is minimized, preventing the network from over-adjusting to uncertain data.
Type-II Fuzzy Sigmoid Activation
To better model uncertainty, we replaced the standard sigmoid with a type-II fuzzy sigmoid activation function. Unlike traditional functions that produce a single "crisp" value, type-II fuzzy sets define a "footprint of uncertainty" (FOU) between upper and lower membership boundaries, mirroring the uncertainty inherent in clinical judgment.
Experimental Results
The model was evaluated using the ISIC 2018 dataset. After data augmentation to balance the "regular" and "irregular" border classes, the dataset comprised 310 images.
The F-MLP achieved an accuracy of 95.2%, significantly higher than the 91.9% achieved by a standard MLP. Comparative analysis showed the F-MLP also outperformed Support Vector Machines (87.1%), Logistic Regression (87.1%), and K-Nearest Neighbors (93.5%). While Random Forests achieved a slightly higher accuracy (96.8%) in some tests, the F-MLP demonstrated superior robustness in handling ambiguous samples and required fewer training iterations to reach peak performance.
Conclusion
The proposed F-MLP provides a robust framework for objectively quantifying skin lesion border irregularity. By explicitly modeling ambiguity through fuzzy gradient descent and type-II activation functions, the network achieves high diagnostic accuracy. While the computational cost of training is higher than standard models, the gain in reliability is significant. Future work will explore "Fuzzy Deep Learning" by integrating these concepts directly into convolutional layers.
Summary updated