The diagnostic problem: Melanoma is the most dangerous form of skin cancer, with over 130,000 new cases annually worldwide. Early detection dramatically improves outcomes - survival rates approach 99% when caught early. Dermoscopy enables non-invasive visual examination, but automated classification systems must handle highly imbalanced datasets and noisy image conditions.
Two complementary approaches: Traditional computer-aided detection (CAD) systems use handcraft features derived from medical algorithms like the ABCD rule (Asymmetry, Borders, Colors, Dermatoscopic Structures). Deep learning CNNs learn features automatically from images but without incorporating medical domain knowledge. This paper proposes fusing both to leverage their complementary strengths.
Mutual information as the fusion engine: Rather than simply concatenating all features, the authors use Mutual Information (MI) - an entropy-based statistical measure - to identify which features carry the most relevant information for the melanoma vs. nevus classification decision. This principled selection discards uninformative features and retains only those that genuinely differentiate the two lesion types.
Validation on ISIC 2018 dataset: The system was tested on the public ISIC 2018 / HAM10000 dataset containing 10,015 dermoscopic images. For binary classification, only melanoma and nevus images were used, with a 75/25 training-test split. The dataset imbalance was corrected using SMOTE oversampling, ensuring fair evaluation on the minority melanoma class.
Preprocessing pipeline: Input dermoscopic images undergo Gaussian filtering to reduce hair and artifact noise, followed by CIEL*a*b* color space transformation. The L (lightness), a* (green-magenta), and b* (blue-yellow) channels are thresholded independently at their mean values and combined via logical intersection to produce a binary mask defining the lesion region of interest (ROI).
Shape features (Asymmetry and Borders): Asymmetry is quantified by rotating the binary mask 180 degrees along its major and minor axes and measuring non-overlapping regions between the original and rotated mask. Additional shape descriptors include area (pixel count), perimeter, circularity (4pi*area/perimeter^2), diameter, and elongation - all grounded in the ABCD rule's A and B components.
Color and texture features: Statistical color features include mean, variance, min, and max values across RGB and CIEL*a*b* channels, representing the ABCD Colors component. Texture features are derived from Gray-Level Co-occurrence Matrices (GLCM), yielding measures like energy, contrast, correlation, variance, entropy, and Inverse Difference Moment (IDM) - capturing the structural patterns emphasized in the Dermatoscopic Structures component.
Deep learning feature extraction: Eight pre-trained CNN architectures were evaluated as generic feature extractors: VGG16 (4,096 features), VGG19 (4,096), ResNet-50 (2,048), Inception v3 (2,048), MobileNet v1 (1,024), MobileNet v2 (1,280), DenseNet-201 (1,920), and Xception (2,048). Each CNN was applied to the ROI image and its penultimate layer activations used as features - not as a classifier, but purely as a learned representation.
The fusion problem: Concatenating handcraft and deep learning features produces a very high-dimensional vector (up to 4,139 features for VGG architectures). Simply using all features increases computation time and risks overfitting. Moreover, not all features contribute equally to distinguishing melanoma from nevus - some CNN activations may carry near-zero discriminative information.
Mutual Information (MI) quantifies feature relevance: MI measures how much information one variable carries about another - in this case, how much each extracted feature contributes to predicting the lesion class. Features with MI values below the mean threshold across all features are discarded; those above are retained. This produces a compact, informative feature subset from both handcraft and deep sources.
High-MI features identified: Analysis of feature rankings showed that color statistics from the b* channel (Mean_b, Min_b, Max_b) and lightness variance (Var_L) had the highest MI values, reflecting the clinical importance of color distribution in distinguishing melanoma from benign lesions. Selected deep learning features were intermixed with these handcraft features, confirming that both types contributed meaningful information.
Classifiers evaluated: After MI-based feature selection, three classifiers were trained: Logistic Regression (LR), Support Vector Machines (SVM with linear and RBF kernels), and Relevant Vector Machines (RVM). The rationale was to test whether the fused features generalize across different classification paradigms, validating that performance was feature-driven rather than classifier-specific.
The imbalance problem in ISIC 2018: The HAM10000 dataset contains significantly more nevus images than melanoma images. When melanoma is the minority class, standard classifiers tend to maximize overall accuracy by over-predicting the majority class, producing systems that achieve good average accuracy but miss the clinically critical melanoma cases.
SMOTE oversampling: Synthetic Minority Oversampling Technique (SMOTE) generates synthetic feature vectors for the minority class by interpolating between existing minority samples using K-nearest neighbors. For each minority sample, SMOTE selects a neighbor and creates a new sample at a random point along the feature-space line connecting them - iterating until both classes are balanced.
New metrics for imbalanced evaluation: Standard accuracy is inadequate for imbalanced datasets. The paper adopts Geometric Mean (G-mean, the square root of sensitivity times specificity), Dominance (sensitivity minus specificity), and Index of Balanced Accuracy (IBA = (1 - Dominance) * G-mean^2). IBA rewards classifiers that perform well on both classes, moderately favoring the minority (melanoma) class without ignoring the majority.
Why balanced evaluation matters clinically: A system that achieves 90% accuracy by correctly classifying all nevi but missing 40% of melanomas is clinically dangerous. The IBA metric, combined with SMOTE, ensures that the reported performance figures reflect true diagnostic utility - a system that reliably identifies melanoma even when nevi are more common.
Comparison across CNN architectures: All eight CNN architectures fused with handcraft features via MI selection outperformed the baseline. MobileNet v2 achieved the best training accuracy of 92.4%, test accuracy of 89.71%, sensitivity of 86.41%, specificity of 90%, AUC of 89.64%, and IBA of 0.80. MobileNet v1 and DenseNet-201 were close runners-up, while VGG16 performed least well (84.9% test accuracy).
Comparison with state-of-the-art methods: Against three published CAD systems, the proposed MobileNet v2 system outperformed on accuracy (92.4% vs. 86.07%, 95%, 85.55% - with the 95% system using a specially preselected dataset). On sensitivity, the proposed method (86.41%) matched published approaches while adding specificity and IBA metrics that were not reported by competitors.
Superiority of MI fusion over concatenation: The key advantage of MI-based feature selection over naive concatenation or PCA reduction is that it explicitly uses information-theoretic relevance to the classification task. Features that carry no predictive signal are discarded regardless of their magnitude, resulting in a more compact and generalizable model.
Robustness of balanced evaluation: The IBA value of 0.80 confirms that the system performs well for both melanoma and nevus classes on the balanced SMOTE dataset - not just for the majority class. The MCC value of 0.795 also indicates strong binary classification performance, close to the maximum of 1.0.
Binary classification limitation: The current framework focuses on melanoma versus nevus - a clinically important but simplified two-class problem. ISIC 2018 contains seven lesion types (actinic keratosis, BCC, dermatofibroma, melanoma, nevus, pigmented benign keratosis, vascular). Extending to multiclass classification requires addressing severe class imbalances across all seven categories.
Dataset dependency: The system was validated only on ISIC 2018. Performance on other dermoscopic datasets, images from different dermatoscope models, different skin tones, or clinical photographs (rather than dermoscopy) remains untested. Cross-dataset validation is essential for demonstrating true generalizability.
Computational efficiency: While MobileNet architectures are designed to be lightweight compared to VGG, the full pipeline (preprocessing, handcraft feature extraction, CNN inference, MI selection, and classification) still requires significant compute for real-time clinical deployment. Optimization for edge devices or mobile dermatoscopes is a natural next step.
Integration with clinical information: The ABCD rule was originally designed to be used alongside patient history (age, lesion evolution, family history). Future work could incorporate EFG features (Evolution, Firmness, Growth) and clinical metadata alongside image features, potentially improving classification accuracy for borderline lesions where the image alone provides insufficient evidence.