Introduction & Objectives

Artificial Intelligence (AI) has shown proficiency in dermoscopy in identifying various types of skin lesions. The segmentation process is crucial for classifying skin lesions, as it enables the model to focus on relevant areas of the image. However, due to the time-consuming nature of manual annotation, the lack of labeled data can significantly impact the performance of segmentation models in real-world applications. Additionally, the presence of black-rimmed lenses in real-life dermoscopic images complicates segmentation, as lenses are weakly represented in labeled data and often lead to mis-segmentation in state-of-the-art models.

We propose a segmentation model that employs a human-in-the-loop strategy and synthetic data to quickly segment new images and enhance performance in multiclass segmentation. The new trained model identifies both the lesion and the lens.

Materials & Methods

The datasets used included ISIC, Derm7PT, and PH2, totaling over 100,000 images, with only 2,594 images annotated for lesion segmentation from the ISIC 2018 dataset. We employed a human-in-the-loop strategy to quickly segment new images and enhanced performance in multiclass segmentation. This strategy involved multiple stages, where each stage trained a new model with an expanding dataset. Initially, the model was trained on the labeled data. The remaining unlabeled data was then segmented using this model. The segmentation results were sorted based on the model’s confidence in its predictions. New masks were human-approved and injected into the training set for the next stage. This iterative process continued until all images were segmented. Inspired by the ABCD rule, the model’s confidence was defined by the shape of the segmented lesion, with the expectation that the lesion found in the image was unique. Additionally, the model included synthetic data to identify lenses. These data were generated from manually extracted lenses from various dermoscopic images and synthetic lenses. The synthetic data were randomly augmented with translation, rotation, color jittering, and affine transformations to enhance the model’s robustness. These augmentations were then randomly applied to the images. The segmentation models utilized a U-Net++ architecture with an EfficientNet-B0 backbone and were evaluated based on Recall, Precision, and DICE score.

Results

On lesion detection, the model achieves a DICE score of 95.9%, Recall of 95.9%, and Precision of 95.8% on real images. The model trained on the entire dataset outperforms the model trained solely on public labeled data, which achieved a DICE score of 81.1%. For lens segmentation, the model achieves a DICE score of 99.6%, Recall of 99.2%, and Precision of 99.8% on synthetic images.

Conclusion

The segmentation of dermoscopic images is a critical step in the classification of skin lesions. Our results show significant improvement in segmentation performance with minimal time investment, thanks to AI-guided human-in-the-loop strategies. Additionally, our findings highlight the feasibility of using synthetic data to develop segmentation models from scratch, offering possibilities for creating effective synthetic datasets to enhance performance of segmentation models for features like hairs without exhaustive labelling. Furthermore, multiclass segmentation models can provide additional valuable information to lesion classification models, such as the type of lens and dermoscop used.

EADV Congress 2025, Paris.