A Patient-Centric Dataset of Images and Metadata for Identifying Melanomas Using Clinical Context (2008.07360v1)

Abstract: Prior skin image datasets have not addressed patient-level information obtained from multiple skin lesions from the same patient. Though artificial intelligence classification algorithms have achieved expert-level performance in controlled studies examining single images, in practice dermatologists base their judgment holistically from multiple lesions on the same patient. The 2020 SIIM-ISIC Melanoma Classification challenge dataset described herein was constructed to address this discrepancy between prior challenges and clinical practice, providing for each image in the dataset an identifier allowing lesions from the same patient to be mapped to one another. This patient-level contextual information is frequently used by clinicians to diagnose melanoma and is especially useful in ruling out false positives in patients with many atypical nevi. The dataset represents 2,056 patients from three continents with an average of 16 lesions per patient, consisting of 33,126 dermoscopic images and 584 histopathologically confirmed melanomas compared with benign melanoma mimickers.

• The paper introduces a novel patient-centric dataset that incorporates clinical context by capturing an average of 16 lesions per patient for improved melanoma detection.
• The methodology features rigorous quality assurance with histopathological validation and a precise 2.7% label removal rate to ensure high data accuracy.
• The dataset’s holistic approach mimics clinical evaluation, enhancing AI tools for robust, real-world melanoma diagnosis and patient management.

A Patient-Centric Dataset of Images and Metadata for Identifying Melanomas Using Clinical Context

This paper details a novel approach towards creating a machine learning dataset that aligns more closely with clinical dermatology practice by incorporating patient-level context. The dataset, introduced by Rotemberg et al., aims to bridge the gap between the single image evaluation paradigm prevalent in AI dermatological studies and the holistic approach used by clinicians, where multiple lesions on a patient are often assessed in tandem.

Summary of the Dataset

The dataset comprises 33,126 dermoscopic images collected from 2,056 patients across several academic and clinical institutions across three continents. A distinguishing aspect is the inclusion of approximately 16 images per patient on average, enabling the contextual assessment of lesions. Among the dataset, there are 584 images of histopathologically confirmed melanomas, juxtaposed with benign mimickers, contributing to a more challenging and clinically relevant training set for algorithm development. This patient-centric approach directly addresses clinical heuristics like the "ugly duckling sign," which considers a lesion's uniqueness against a patient’s typical lesion pattern.

Methodological Considerations

The dataset was meticulously curated through a series of quality assurance steps to ensure high label accuracy and image quality. Experts employed a proprietary software tool, ‘Tagger,’ to facilitate image verification, flagging erroneous labels with a removal rate of 2.7%. The lesions were histopathologically validated when possible, and non-biopsied lesions were monitored for at least six months to be deemed benign. The datasets were drawn from high-risk clinics with robust imaging practices, ensuring a comprehensive collation of dermoscopy modalities, including polarized and non-polarized, contact, and non-contact images.

Implications and Future Directions

The primary implication of this work is its potential to enhance the fidelity of AI systems in dermatology, moving beyond isolated lesion assessment to a more holistic view resembling clinical practice. By offering a framework that can evaluate lesions in their biological context, this dataset may help algorithms mimicking the diagnostic acumen of dermatologists, especially in complex cases involving patients with a multitude of atypical nevi.

Moreover, the integration of such datasets can facilitate the development of AI tools that might assist clinicians, especially in resource-limited settings where specialist access is restricted. These tools could significantly aid in early detection and management of melanoma, a leading cause of cancer mortality.

While the dataset presents significant advancements, it also raises considerations for future research directions. Continued efforts should focus on expanding the dataset to include diverse skin types, acknowledging the current challenges and bias toward non-representative populations. Additionally, future studies may explore the incorporation of sequential imaging over time, offering dynamic insights into lesion evolution.

Conclusion

This paper introduces a pioneering dataset that reflects clinical dermatology's contextual evaluation strategy. By providing comprehensive lesion context, it promises to elevate the efficacy and applicability of AI models in real-world scenarios. As the field progresses, enhancements in dataset inclusivity and temporal scope are anticipated to further enrich AI's diagnostic capabilities.