Problem

The skin, as the human body's largest organ, is susceptible to skin cancer, often originating in the epidermal layer due to factors like UV radiation. Early detection is paramount for successful treatment and improved survival rates. However, automated classification of skin lesions presents significant challenges due to the visual similarities between various lesion types and the need for fine-grained contextual information. Traditional machine learning methods struggle with a scarcity of annotated medical image datasets, considerable variability in lesion appearance, and the need for labor-intensive manual feature engineering. While deep learning, particularly Convolutional Neural Networks (CNNs) and transfer learning, has emerged as a promising avenue, accurately classifying diverse skin diseases remains an ongoing and complex problem that necessitates more robust and efficient solutions.

Method

The proposed methodology for early skin cancer classification employs a sophisticated pipeline centered on a soft voting ensemble of Convolutional Neural Networks. The process begins with extensive preprocessing of three benchmark datasets: HAM10000, ISIC 2016, and ISIC 2019. This stage includes rebalancing the datasets using inverse class weights to counteract class imbalance, normalizing image pixel values to a [0,1] range, augmenting minority classes with affine transformations (such as rotation, zooming, cropping, flipping, and translation), and applying a sequence of filtering techniques (Gaussian blur, median filter, Sobel edge detection, and histogram equalization) to reduce noise and enhance relevant features. Subsequently, a hybrid dual encoder, leveraging transfer learning, performs image segmentation. This model utilizes two encoder instances that combine their outputs to capture both fine-grained and global image features, thereby precisely segmenting lesion areas. For the final classification, a soft voting ensemble is constructed from three pre-trained CNN architectures: MobileNetV2, VGG19, and InceptionV3. MobileNetV2 is selected for its efficiency and suitability for resource-constrained environments, VGG19 for its deep architecture capable of abstracting high-level features, and InceptionV3 for its multi-scale feature extraction capabilities. The ensemble combines the weighted predictions from these models to achieve a balance between diagnostic accuracy and inference speed, aiming for a versatile and robust system for real-world medical deployment.

Results

The experimental evaluation, conducted on an NVIDIA Tesla T4 GPU, demonstrated that the proposed soft-voting ensemble model significantly outperformed individual MobileNetV2, VGG19, and InceptionV3 models across key metrics, particularly recall, which is crucial for medical diagnostic systems to avoid false negatives. The ensemble achieved impressive lesion recognition accuracies of 96.32% on the HAM10000 dataset, 90.86% on the ISIC 2016 dataset, and 93.92% on the ISIC 2019 dataset. While the ensemble approach naturally incurred a slightly longer processing time compared to single CNN models, it presented the optimal trade-off, delivering high accuracy within a reasonable inference time. Analysis of Area Under the Curve (AUC) values revealed varying degrees of distinguishability for different lesion types, with high AUCs for Nevus indicating reliable differentiation, and lower AUCs for melanoma, dermatofibroma, vascular, and BCC suggesting more challenging classifications. Furthermore, a comparative analysis against several contemporary advanced classification methods highlighted the superior and consistent performance of the proposed ensemble system across all tested datasets, affirming its robustness and its ability to maintain high recall rates for critical cancerous lesion detection.

Implications

This research offers a promising advancement in the field of automated skin lesion classification, presenting a method that effectively balances diagnostic accuracy with operational speed through a soft voting ensemble of CNNs combined with robust preprocessing and segmentation. The implications of this work are significant for real-world dermatological applications, enabling more efficient and reliable early detection of skin cancer. The system's design, incorporating lightweight models like MobileNetV2, facilitates its potential deployment on mobile and edge devices, thereby enhancing accessibility to real-time screening, especially in underserved or remote areas. By prioritizing high recall rates, the system minimizes the risk of overlooking critical cancerous lesions, thus contributing to improved patient safety and outcomes. Although the model demonstrates strong performance and surpasses existing approaches, the authors acknowledge that there is still room for further improvement in its classification accuracy. Future endeavors will focus on refining the model to enhance its overall reliability and efficiency, particularly to mitigate the severe consequences of misclassifying cancerous skin lesions.