TLDR: This research introduces a novel Deep Convolutional Neural Network (DCNN) with custom PReLU layers and Explainable AI (LIME) for highly accurate and interpretable detection of the “bluish veil” (BWV) in dermoscopic images, a critical melanoma indicator. By creating annotated datasets and outperforming existing methods, the model provides a robust tool for early melanoma diagnosis, with LIME offering insights into its decision-making process.
Melanoma, one of the deadliest forms of skin cancer, accounts for thousands of fatalities globally each year. Early and accurate diagnosis is paramount for improving patient outcomes, as melanoma is highly treatable when detected in its initial stages before it spreads to other parts of the body. Among the various clinical methods developed to aid in early diagnosis, the presence of a bluish, blue-whitish, or blue-white veil (BWV) is a particularly critical feature for identifying melanoma. This veil is characterized by a blue and structureless zone with an overlying white “ground-glass” haze, often found in raised areas of a lesion.
Despite its significant clinical importance, detecting the BWV can be challenging due to its subtle appearance. The veil often presents as a faint mix of gray, blue, and white patches, which can be easily overlooked against the complex background of other lesion colors. This subtlety makes accurate detection a task prone to human error, especially in busy clinical environments. Current manual examination techniques, such as the Three-Point Checklist and Seven-Point Checklist, rely heavily on visual assessment and may not be sufficiently robust for consistent BWV identification.
While Deep Convolutional Neural Networks (DCNNs) have shown immense promise in various image recognition tasks, including medical imaging, their specific application to BWV detection in dermoscopic images has been limited. Existing methods often do not adequately address the complexities involved in precisely identifying this feature, highlighting a clear need for advanced machine learning techniques to enhance detection accuracy and reliability.
This study introduces a novel approach to automate the detection of BWV in dermoscopic images by leveraging DCNNs combined with Explainable Artificial Intelligence (XAI) techniques. Unlike previous studies that often rely on traditional image processing and manual inspection, this approach automates the detection process, aiming to enhance both accuracy and consistency. The integration of XAI further provides crucial insights into the model’s decision-making process, fostering transparency and trust in its predictions.
A key objective of this research was to create an annotated dataset of skin lesions, as many existing datasets lack specific annotations for BWV. This was achieved using a proposed imaging algorithm based on color threshold techniques. This algorithm converts non-annotated skin lesion datasets into annotated ones by identifying specific RGB color ranges associated with BWV, mimicking human perception of these subtle colors. The process involved systematically analyzing each image by dividing it into small 16×16 pixel patches and scrutinizing these patches to identify colors within the established RGB channel range for BWV. This ground truth establishment was validated by clinical experts, ensuring its reliability.
The researchers designed and trained a custom 31-layer DCNN, inspired by existing network architectures but fine-tuned with various ablation studies to optimize its parameters. Instead of standard activation functions, custom layers utilizing the Parametric Rectified Linear Unit (PReLU) were incorporated. PReLU is an advanced activation function that introduces a learnable parameter for negative inputs, allowing the network to adaptively learn the optimal slope for these values, thereby enhancing adaptability and improving the model’s efficiency and accuracy compared to simpler functions like ReLU or Leaky ReLU.
The model was trained separately and collectively on three distinct dermoscopic datasets: PH2, ISIC archive, and Derm7pt. To address the limited number of images in some datasets, data augmentation techniques—including image rotation, flipping, and zooming—were applied to increase the training data volume, which is essential for robust deep learning model training.
The performance of the proposed DCNN was rigorously evaluated using standard metrics such as accuracy, precision, sensitivity, F1-score, and specificity. The model demonstrated superior performance compared to conventional BWV detection models across different datasets. For instance, the PReLU-based DCNN achieved a testing accuracy of 85.71% on the augmented PH2 dataset, 95.00% on the augmented ISIC archive dataset, 95.05% on the combined augmented (PH2+ISIC archive) dataset, and 90.00% on the Derm7pt dataset. In some experiments, the PReLU-based DCNN achieved remarkable results, including 100% precision and 100% specificity, or 100% sensitivity, indicating a very low rate of false positives or false negatives, respectively.
To provide transparency and build trust in the model’s predictions, the Local Interpretable Model-agnostic Explanations (LIME) technique was employed. LIME helps interpret the DCNN’s decision-making process by highlighting the specific regions in the input images that most influenced the model’s classification. This analysis revealed that the network primarily focused on the red, brown, and bluish regions of the lesions for its BWV predictions. Specifically, the bluish and whitish areas were identified as key indicators for the BWV class, while red, brown, and light-brown colors indicated the non-BWV class. This interpretability is crucial for dermatologists and skin cancer experts to understand and trust the AI’s recommendations.
Also Read:
- CorrDetail: Enhancing Deepfake Detection with Self-Correction and Visual Detail
- Evaluating Explainable AI for Remote Sensing Imagery: A Comprehensive Analysis
This innovative approach, detailed further in the research paper, significantly improves the detection of BWV in skin lesions, providing a robust and interpretable tool for early melanoma diagnosis. Future work will involve extending this methodology to identify other prevalent colors and patterns in dermoscopic images, with the ultimate goal of constructing an integrated system to facilitate the timely detection of melanoma.
