Source-linked AI summary
Skin Lesion Analyser: An Efficient Seven-Way Multi-Class Skin Cancer Classification Using MobileNet
Saket S. Chaturvedi, Kajol Gupta, Prakash. S. Prasad
TL;DR
Skin cancer screening is challenged by disease burden, visually similar lesions, and variable diagnostic accuracy. This paper uses transfer learning with an ImageNet-pretrained MobileNet trained on HAM10000 for seven-way dermoscopic classification, achieving 83.1% overall accuracy with 91.36% top2 and 95.34% top3 accuracy. The authors conclude that the approach matched expert-dermatologist performance across seven diagnostic tasks, while future work should incorporate personalized patient data.
Problem
Skin cancer is a major health problem, and dermoscopic diagnosis can be less accurate for inexperienced dermatologists; prior automated methods lacked generality and multi-class performance.
Method
The study fine-tuned an ImageNet-pretrained MobileNet on HAM10000 dermoscopy images, using transfer learning and augmentation of minority classes.
Results
83.1% overall accuracy, 91.36% top2 accuracy, and 95.34% top3 accuracy were achieved for seven classes, with weighted precision, recall, and F1-score of 0.89, 0.83, and 0.83.
Takeaways & Limitations
The authors report that the seven-way MobileNet method matched expert dermatologists’ performance across seven diagnostic tasks.
Takeaways & Limitations
Future work may add personalized data such as genes, age, and color to develop personalized computer-aided diagnosis systems.
Abstract
from arXiv · showhide
Skin cancer, a major form of cancer, is a critical public health problem with 123,000 newly diagnosed melanoma cases and between 2 and 3 million non-melanoma cases worldwide each year. The leading cause of skin cancer is high exposure of skin cells to UV radiation, which can damage the DNA inside skin cells leading to uncontrolled growth of skin cells. Skin cancer is primarily diagnosed visually employing clinical screening, a biopsy, dermoscopic analysis, and histopathological examination. It has been demonstrated that the dermoscopic analysis in the hands of inexperienced dermatologists may cause a reduction in diagnostic accuracy. Early detection and screening of skin cancer have the potential to reduce mortality and morbidity. Previous studies have shown Deep Learning ability to perform better than human experts in several visual recognition tasks. In this paper, we propose an efficient seven-way automated multi-class skin cancer classification system having performance comparable with expert dermatologists. We used a pretrained MobileNet model to train over HAM10000 dataset using transfer learning. The model classifies skin lesion image with a categorical accuracy of 83.1 percent, top2 accuracy of 91.36 percent and top3 accuracy of 95.34 percent. The weighted average of precision, recall, and f1-score were found to be 0.89, 0.83, and 0.83 respectively. The model has been deployed as a web application for public use at (https://saketchaturvedi.github.io). This fast, expansible method holds the potential for substantial clinical impact, including broadening the scope of primary care practice and augmenting clinical decision-making for dermatology specialists.
1 Introduction
Skin cancer diagnosis is important because incidence is linked to UV exposure, while visual and dermoscopic assessment can vary with lesion similarity and clinician experience. The study addresses limited generality and multi-class performance by applying transfer learning with MobileNet to seven dermoscopic cancer classes.
- UV radiation can damage skin-cell DNA and trigger uncontrolled growth associated with skin cancer.
- 99 percent five-year melanoma survival is reported when diagnosis and screening occur at an early stage.
- 80 percent diagnostic accuracy was reported for dermatologists with more than ten years’ experience, versus 62 percent for those with 3–5 years.
- Deep learning has supported automated skin-cancer screening and has matched or exceeded dermatologists in a 129,450-image CNN benchmark.
- Previous dermoscopic classification work lacked generality and produced unsatisfactory multi-class results.
- The study fine-tuned ImageNet-pretrained MobileNet on 10,015 HAM10000 dermoscopy images for seven-class classification.
2 Method
The method combines HAM10000 dermoscopy data, preprocessing and augmentation, and an ImageNet-pretrained MobileNet architecture. Class imbalance is addressed by augmenting minority classes, while several accuracy and precision–recall metrics evaluate performance.
- 2.1 Dataset: HAM10000 contains 10,015 dermoscopy images across seven lesion classes, with class counts ranging from 115 Dermatofibroma to 6,705 Melanocytic nevi.
- 2.2 Data pre-processing: The dataset was downscaled from 600X450 to 224X224 pixels and split into 9,077 training and 938 validation images without duplication in training data.
- 2.3 Data augmentation: Data augmentation generated approximately 6,000 images in each minority class, producing 38,569 training images.
- 2.4 Model: MobileNet uses a lightweight architecture and was pretrained on 12,80,000 images containing 1,000 ImageNet object classes.
- 2.5 Evaluation metrics: Accuracy, Top2 accuracy, Top3 accuracy, micro averages, and weighted precision, recall, and F1-score were used for evaluation.
3 Results
The study evaluates MobileNet on HAM10000 using classification metrics, confusion-matrix analysis, exploratory dataset analysis, and training-validation curves. Results include seven-class performance measures, class-specific prediction differences, and a web deployment.
- Data-set analysis: The HAM10000 analysis identified a peak skin-cancer age of 45 years, greater prevalence between ages 30 and 70, and commonly affected body regions.
- Model performance: Weighted-average precision, recall, and F1-score were 89%, 83%, and 83%, respectively, on 938 unknown validation images.
- Model performance: 83.15 percent categorical accuracy, 91.36 percent top2 accuracy, 95.3 percent top3 accuracy, and 83 percent recall were achieved using MobileNet.
- Comparison with previous work: The seven-class comparison reports prior accuracies of approximately 66 percent to 81 percent and recalls of 60 percent to 76 percent, while this study reports higher accuracy and recall.
- Confusion matrix: 696 of 751 Melanocytic nevi images were correctly classified, compared with 26 of 30 Basal cell carcinoma and 27 of 39 Melanoma images.
- Confusion matrix: Benign keratosis was most challenging, with only ten correct predictions, because of similar appearance to Melanoma and Melanocytic nevi.
- Loss and accuracy curves: A small gap between training and validation curves was interpreted as a good fit and generalization to unknown images.
- Deployment: The authors developed a web application for automated online classification of dermoscopy skin-lesion images.
4 Conclusion
The study presents MobileNet as an efficient approach for automated seven-class dermoscopic skin-cancer classification. It reports expert-comparable performance and identifies personalized patient data as a direction for future work.
- MobileNet achieved 83.1% overall accuracy, 91.36% top2 accuracy, and 95.34% top3 accuracy across seven diagnostic classes.
- Weighted-average precision, recall, and F1-score were 89%, 83%, and 83%, respectively.
- The authors conclude that MobileNet can support an efficient real-time computer-aided system for automated medical diagnosis.
- Future work may add personalized data such as genes, age, and color for skin-cancer diagnosis.