Design of a Machine Learning System to Predict the Thickness of a Melanoma Lesion in a Non-Invasive Way from Dermoscopic Images
1
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Objectives: Melanoma is the deadliest form of skin cancer, but it can be fully cured through early detection and treatment in99% of cases. Our aim was to develop a non-invasive machine learning system that can predict the thickness of a melanomalesion, which is a proxy for tumor progression, through dermoscopic images. This method can serve as a valuable tool inidentifying urgent cases for treatment. Methods: A modern convolutional neural network architecture (EfficientNet) wasused to construct a model capable of classifying dermoscopic images of melanoma lesions into three distinct categories basedon thickness. We incorporated techniques to reduce the impact of an imbalanced training dataset, enhanced the generalizationcapacity of the model through image augmentation, and utilized five-fold cross-validation to produce more reliablemetrics. Results: Our method achieved 71% balanced accuracy for three-way classification when trained on a small publicdataset of 247 melanoma images. We also presented performance projections for larger training datasets. Conclusions: Ourmodel represents a new state-of-the-art method for classifying melanoma thicknesses. Performance can be further optimizedby expanding training datasets and utilizing model ensembles. We have shown that earlier claims of higher performance weremistaken due to data leakage during the evaluation process.
Abstract
Objectives: Melanoma is the deadliest form of skin cancer, but it can be fully cured through early detection and treatment in99% of cases. Our aim was to develop a non-invasive machine learning system that can predict the thickness of a melanomalesion, which is a proxy for tumor progression, through dermoscopic images. This method can serve as a valuable tool inidentifying urgent cases for treatment. Methods: A modern convolutional neural network architecture (EfficientNet) wasused to construct a model capable of classifying dermoscopic images of melanoma lesions into three distinct categories basedon thickness. We incorporated techniques to reduce the impact of an imbalanced training dataset, enhanced the generalizationcapacity of the model through image augmentation, and utilized five-fold cross-validation to produce more reliablemetrics. Results: Our method achieved 71% balanced accuracy for three-way classification when trained on a small publicdataset of 247 melanoma images. We also presented performance projections for larger training datasets. Conclusions: Ourmodel represents a new state-of-the-art method for classifying melanoma thicknesses. Performance can be further optimizedby expanding training datasets and utilizing model ensembles. We have shown that earlier claims of higher performance weremistaken due to data leakage during the evaluation process.
참고문헌 (18)
- "SEER Cancer Statistics Review (CSR) 1975-2018" (2021)[URL]
- LeCun Y, "Deep learning", Nature, 521(7553), pp.436~444 (2015)[DOI]
- Tschandl P, "Human-computer collaboration for skin cancer recognition", Nat Med, 26(8), pp.1229~1234 (2020)[DOI]
- Tschandl P, "Comparison of the accuracy of human readers versus machine-learning algorithms for pigmented skin lesion classification : an open, webbased, international, diagnostic study", Lancet Oncol, 20(7), pp.938~947 (2019)[DOI]
- Rotemberg V, "The role of public challenges and data sets towards algorithm development, trust, and use in clinical practice", Semin Cutan Med Surg, 38(1), pp.E38~E42 (2019)[DOI]
- Rotemberg V, "A patient-centric dataset of images and metadata for identifying melanomas using clinical context", Sci Data, 8(1), pp.34~ (2021)[DOI]
- Shaikh WR, "melanoma thickness and survival trends in the United States, 1989 to 2009", J Natl Cancer Inst, 108(1), pp.djv294~ (2015)[DOI]
- Rubegni P, "Evaluation of cutaneous melanoma thickness by digital dermoscopy analysis : a retrospective study", Melanoma Res, 20(3), pp.212~217 (2010)[DOI]
- Saez A, "Machine learning methods for binary and multiclass classification of melanoma thickness from dermoscopic images", IEEE Trans Med Imaging, 35(4), pp.1036~1045 (2016)[DOI]
- Sanchez-Monedero J, "Classification of melanoma presence and thickness based on computational image analysis", Proceedings of the 11th International Conference on Hybrid Artificial Intelligent Systems (HAIS), pp.427~438 (2016)
- Jaworek-Korjakowska J, "Melanoma thickness prediction based on convolutional neural network with VGG-19 model transfer learning", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pp.2748~2756 (2019)
- Chawla NV, "SMOTE : synthetic minority over-sampling technique", J Artif Intell Res, 16, pp.321~357 (2002)[DOI]
- Kawahara J, "7-Point checklist and skin lesion classification using multi-task multi-modal neural nets", IEEE J Biomed Health Inform, 23(2), pp.538~546 (2018)[DOI]
- Tan M, "EfficientNet : rethinking model scaling for convolutional neural networks", Proc Mach Learn Res, 97, pp.6105~6114 (2019)
- Kuhn M, "Feature engineering and selection:a practical approach for predictive models" (2019), CRC Press
- Somfai E, "Minimizing false negative rate in melanoma detection and providing insight into the causes of classification" (2021)
- "Data preparation for machine learning:data cleaning, feature selection, and data transforms in Py" (2020)[URL]
- Frid-Adar M, "Synthetic data augmentation using GAN for improved liver lesion classification", Proceedings of 2018IEEE 15th International Symposium on Biomedical Imaging (ISBI), pp.289~293 (2018)