Abstract
Skin cancer remains a formidable global health challenge, necessitating precise and timely diagnostic methodologies. This study focuses on advancing the field through the development and evaluation of deep learning algorithms tailored for skin cancer detection using 3D Total Body Photography (3D-TBP). Leveraging the ISIC 2024 dataset, comprising a diverse array of high-resolution skin lesion images, our approach integrates rigorous data preprocessing, sophisticated model architecture design, and meticulous performance evaluation.
The dataset underwent meticulous curation and augmentation to bolster model robustness and generalizability. A specialized convolutional neural network (CNN) architecture was crafted, specifically optimized for analysing single-lesion crops extracted from 3D-TBP images. This CNN framework leverages transfer learning, combining efficient feature extraction with finely tuned classification layers to maximize diagnostic accuracy.
Training was conducted on a high-performance computing platform, employing advanced techniques such as batch normalization and dropout regularization to mitigate overfitting and enhance model generalization. Hyperparameter tuning and cross-validation protocols were rigorously implemented to ensure optimal model configuration and validation.
Evaluation metrics were cantered on the partial area under the ROC curve (pAUC) with a focus on achieving an 80% true positive rate (TPR), aligning closely with competition benchmarks and clinical diagnostic requirements. Our developed CNN model demonstrated robust performance during validation, surpassing an impressive pAUC of 85% on the test dataset. Notably, the model exhibited superior discriminatory abilities across various skin types and lesion morphologies, effectively distinguishing between malignant and benign lesions.
In conclusion, this study presents a cutting-edge AI-driven approach for skin cancer detection using 3D-TBP, showcasing substantial advancements in automated dermatological diagnosis. The findings underscore the potential of AI technologies to revolutionize clinical practice, offering enhanced diagnostic precision and efficiency. This research paves the way for further exploration and deployment of AI-driven solutions in dermatology, aiming to improve patient outcomes and streamline healthcare delivery.