Ensemble based skin cancer classification and it’s real-time implementation

Abstract

Efficient classification of skin cancer plays a crucial role in early detection and treatment. Deep learning algorithms have emerged as a promising approach for accurate skin cancer subtype prediction by leveraging intricate patterns and correlations from large clinical datasets. This thesis proposes two novel ensemble techniques: one based on the Levy Stable probability density function (PDF) and another based on the Beta function ensemble technique. The Levy Stable model employs a robust ranking mechanism and is optimized for real-time hardware implementation. Five standard convolutional neural network (CNN) models, including DenseNet-121, EfficientNet-B3, MobileNet-V3-Large, Inception-V3, and ResNet-101, are utilized along with transfer learning to generate confidence scores that enhance classification accuracy. The ensemble technique combines these scores using a rank-based fusion approach based on the Beta function, surpassing traditional ensemble methods by adapting the priority based on classifier confidence scores for each instance. The proposed Beta ensemble technique achieves an F1-score of 83% and an Accuracy of 91%. The effectiveness of the proposed techniques is evaluated on the HAM10k dataset, demonstrating superior performance compared to existing approaches. Additionally, optimization for real-time hardware implementations is conducted using the OpenVINO toolkit, achieving high-performance inference rates on different hardware configurations. Specifically, the Levy Stable PDF achieves 110 FPS and 91 FPS for float 16 and float 32 precision, respectively, on an Intel i7 CPU, while the Beta function achieves 33.56 FPS and 26.80 FPS for float16 and float32 precision, respectively, on an Intel iRIS Xe GPU