TL;DR
This paper introduces the SLICE-3D dataset, a massive collection of over 400,000 skin lesion images extracted from 3D Total Body Photographs (TBP). This dataset aims to overcome limitations of existing datasets by providing images with quality comparable to smartphone photos, enabling the development of AI models for skin cancer detection in real-world, non-specialist settings.
What Problem Does This Paper Address?
Current AI models for skin cancer detection primarily rely on dermoscopic images, which are high-quality but require specialized equipment and trained clinicians. This creates a bottleneck, limiting the accessibility of advanced AI tools to primary care or remote settings where such equipment isn't available. Existing datasets also suffer from selection bias, often underrepresenting common benign lesions, which can lead to AI models that perform poorly on the general population. The SLICE-3D dataset addresses these issues by providing a large, diverse, and standardized collection of skin lesion images that mimic the quality of images captured by non-experts, paving the way for more broadly applicable AI diagnostic tools.
Methods
The core of this work is the creation and release of the SLICE-3D dataset. Images were extracted from 3D Total Body Photographs (TBP) captured using the VECTRA WB360 imaging system. This system uses multiple cameras to capture the entire body surface, and proprietary software (Lesion Visualizer) automatically detects and tags lesions. The "SLICE-3D" dataset specifically refers to cropped "tile" images of individual lesions, standardized to a 15mm x 15mm field of view. These tiles are comparable in resolution to smartphone images. The dataset includes both "strong-labels" (lesions with confirmed pathology reports within 3 months of imaging) and "weak-labels" (lesions assumed benign based on clinical evaluation and lack of biopsy). A rigorous quality assurance process was implemented, including manual review to remove PHI and non-lesion images, and duplicate detection.
Datasets
The primary contribution is the SLICE-3D dataset, containing over 400,000 skin lesion image crops. This dataset is distinct from existing dermoscopic image datasets due to its image quality (smartphone-like), origin (3D TBP), and the inclusion of a wider range of lesions, including common benign ones. The paper also mentions the ISIC2024 Tile Export Tool as a key component for data extraction.
Key Insights and Ideas
- Bridging the Gap: The SLICE-3D dataset is a crucial step towards making AI-powered skin cancer detection accessible beyond specialized dermatology clinics. By using images similar to those captured by smartphones, it opens doors for AI tools in primary care, telehealth, and even direct-to-patient applications.
- Addressing Selection Bias: The use of 3D TBP, which captures the entire body, helps to mitigate the selection bias found in datasets composed solely of dermoscopic images. This means the dataset is more representative of the spectrum of skin lesions encountered in real-world scenarios, including common benign lesions that are often underrepresented.
- Standardization for Robustness: While mimicking smartphone image quality, the dataset maintains a degree of standardization in terms of field of view and extraction process. This is vital for training AI models that can generalize well and are not overly sensitive to variations in lighting or capture conditions.
- "Weak-labeling" for Real-World Relevance: The inclusion of "weak-labels" for benign lesions acknowledges that many skin lesions are not biopsied. This approach allows for the training of models that can differentiate between concerning and non-concerning lesions even when definitive pathological confirmation is absent, reflecting typical clinical workflows.
Key Results
The paper's primary result is the creation and release of the SLICE-3D dataset, comprising over 400,000 de-identified, diagnostically labeled skin lesion images. The dataset was collected from seven dermatologic centers worldwide. The paper details the data collection and curation process, including the sub-selection algorithm for "strong" and "weak" labels. While the paper focuses on dataset creation, it implies that this dataset will be instrumental for developing AI algorithms that can improve clinical workflows and early skin cancer detection.
Limitations and Future Work
- Skin Tone: The authors acknowledge that it was not possible to document the skin tone of individuals in this dataset due to the retrospective nature of the collection. This is a limitation for developing AI models that are equitable across all skin types.
- Image Quality Variability: While aiming for smartphone-like quality, the pixel resolution of the tile images can vary due to the capture apparatus. Users are advised to use relative pixel distance for analysis.
- Future Work: The paper suggests that this dataset can be used to build upon previous applications of skin imaging for cancer detection. The SLICE-3D dataset was used as the training dataset for the ISIC 2024 Grand Challenge, indicating its immediate utility for advancing AI in this field. Future work will likely involve developing and evaluating AI models trained on this dataset for various skin cancer detection and classification tasks.
Summary updated