TLDR: This research introduces a novel two-part AI-driven method for detecting asymmetric skin lesions from dermoscopic images. It combines a geometric pattern analysis (GSAA) to objectively classify lesion shapes and generate data, with a powerful CNN-SVM model for highly accurate classification of asymmetric, half-symmetric, and symmetric lesions, outperforming previous methods and aiding early melanoma diagnosis.
The shape of a skin lesion holds crucial clues for diagnosing various skin conditions, especially in the early detection of melanoma, a serious form of skin cancer. Dermatologists often rely on visual cues like asymmetry, border irregularity, color variation, diameter, and evolution (the ABCDE rule) to assess moles and lesions. Among these, asymmetry—where one half of a lesion doesn’t match the other—is a vital indicator of potential malignancy.
Understanding Skin Lesion Asymmetry
Traditionally, assessing lesion asymmetry can be subjective, leading to variations in expert opinions. This challenge highlights the need for more objective and automated methods. While artificial intelligence has made significant strides in medical imaging, specifically in distinguishing between benign and malignant lesions, there has been less focus on automatically identifying asymmetric skin lesions. Furthermore, deep learning models, which are highly effective, require large amounts of labeled data, and datasets specifically annotated for lesion asymmetry are scarce.
The Two-Pronged Approach
To address these gaps, a new research paper introduces an innovative two-part approach for detecting asymmetric skin lesions. The first part proposes a novel image processing algorithm based on geometric patterns, designed to help both experts and non-experts understand lesion asymmetry and to generate much-needed labeled data. The second part utilizes a sophisticated deep learning model to classify lesion shapes.
Geometric Shape-Based Asymmetry Analysis (GSAA)
The first component, called Geometry Shape-Based Asymmetry Analysis (GSAA), works by taking a binary image of a lesion (where the lesion area is clearly defined). It then divides this image into four sections using two perpendicular axes that cross at the lesion’s center. The algorithm counts the number of white pixels (representing the lesion area) in each of these four sections. By comparing the ratios of these pixel counts, the GSAA determines if the lesion is Symmetric, Half-Symmetric (symmetric along one axis), or Asymmetric. For instance, if the ratio of pixel counts between two opposing sections falls within a specific range (0.90 to 1.10), they are considered symmetrical. This method proved highly accurate, achieving a 99.00% detection rate for dermatological asymmetric lesions in one dataset (PH2) and 99.06% in another (ISIC2016). This geometric approach is particularly valuable for creating annotated datasets from previously unlabeled images, providing a reliable ‘ground truth’ for training advanced AI models.
Leveraging Deep Learning with CNN-SVM
The second part of the research involves a powerful combination of a Convolutional Neural Network (CNN) and a Support Vector Machine (SVM) classifier. Pre-trained CNN models, specifically ResNet18, ResNet50, and ResNet101, are employed to extract intricate features like shape, color, and texture from the dermoscopic images. These extracted features are then fed into a multiclass SVM classifier, which is trained to categorize the lesions into three distinct classes: Asymmetric, Half-Symmetric, and Symmetric. This hybrid CNN-SVM approach leverages the CNN’s ability to learn complex visual patterns and the SVM’s strength in classification.
Also Read:
- Advancing Breast Cancer Screening with AI: A New Method for Generating Paired Medical Images
- NuSeC: A New Resource for AI-Powered Breast Cancer Nuclei Segmentation
Impressive Results and Future Outlook
The CNN-SVM models demonstrated exceptional performance in classifying lesion shapes. In experiments, the best performance was achieved with the ResNet101+SVM combination, showing a 94% Kappa Score, 95% Macro F1-score, and 97% Weighted F1-score. This proposed method consistently outperformed existing state-of-the-art techniques in lesion asymmetry analysis, even when trained on one dataset and tested on another, highlighting its robustness and accuracy. The research indicates that the ResNet101+SVM model is particularly well-suited for practical application in dermatology for assessing skin lesion asymmetry.
This study significantly advances the field of dermatological research by providing a reliable, accurate, and consistent automated method for analyzing skin lesion shapes. By reducing the burden on dermatologists and improving the precision of analysis, this technology can lead to earlier diagnoses and better patient outcomes for conditions like melanoma. For more details, you can refer to the full research paper here.
Future work aims to refine the geometric analysis by splitting images into more parts (e.g., eight parts by 45-degree angles) and to separately analyze other lesion properties like border, structure, and color, combining their individual scores for an even more comprehensive assessment.
