← ResourcesPublications
Artificial intelligence quantifies key prognostic factors of melanoma from H&E images
01 SEPT 2024
Abstract
Background & objectives
Breslow Thickness and Ulceration are key to the assessment of Melanoma survival and relapse outcomes. We introduce and compare two approaches based on deep neural networks to quantify and assess these factors: a fully and weakly supervised approaches.
Methods
In this study, we use two datasets (n=350 and 540) and investigated two learning approaches. First, a segmentation model is trained with pixel-wise annotations. A second network is trained to learn from image-level labels. The performance in predicting Breslow thickness and identifying the presence of ulceration in Melanoma H&E-stained whole slide images is assessed on a hold-out dataset.
Results
We show that a network can learn the representation of these pronostic factors without any annotation, precision of 0.85 and 0.92 for Breslow and Ulceration compared to 0.81 and 0.89 with fully-supervised network.
Conclusion
Training models without any pixel-wise annotation outperforms annotation-based learning. Its superior performance, combined with its practical advantages in terms of training requirements, suggests that it could serve as a valuable asset in the fight against melanoma, subject to further validation and integration into clinical workflows. Furthermore, we believe that this study lays a foundation for future AI systems to prognosticate not just melanoma severity but also potential patient survival duration and cancer recurrence probabilities.