Proscia, a provider of digital and computational pathology solutions, has announced new research on improving the generalisation of an artificial intelligence (AI) classification model with contrastive self-supervised learning.
The results, which include a 56.3% increase in melanoma detection sensitivity, will be presented at the Conference on Neural Information Processing Systems (NeurIPS) 2022.
Proscia’s retrospective study, “Learning SimCLR Representations for Improving Melanoma Whole Slide Images Classification Model Generalization” was conducted with data from three sites. The study investigated the impact of extended training time and different augmentations on an AI model that leverages the SimCLR framework of contrastive self-supervised learning to classify melanoma cases.
The results show that optimising these factors can improve the generalisation of an AI model trained with data from two sites, demonstrating a 56.3% increase in melanoma detection sensitivity when evaluated on images from the third site. As melanoma is the deadliest form of skin cancer and often challenging to diagnose, the findings illustrate the promise of contrastive self-supervised learning to help lower the misdiagnosis rate.
“Very often, if you can’t build AI that generalises to new sites, you can’t build AI that makes an impact,” said Julianna Ianni, PhD, Proscia’s VP of AI Research and Development. “Beyond potentially improving melanoma detection, our work highlights the importance of key aspects of training AI that often go overlooked.”
AI development in pathology has been limited by the difficulty of obtaining annotated images to train models. Self-supervised learning has emerged as an approach for overcoming this challenge by leveraging unannotated data. While it has resulted in performance improvements for some models, they often fail to generalise to new sites due to the variability of data in routine practice. Contrastive learning, when applied to self-supervised learning, has yielded promising results for model generalisation as indicated by Proscia’s study. Such results show its potential to increase the volume of images available for training highly-generalisable models.
Details on Proscia’s poster presentation are as follows:
Title: “Learning SimCLR Representations for Improving Melanoma Whole Slide Images Classification Model Generalization”
Event: Conference on Neural Information Processing Systems
Date: 2 December, 2022 from 10:30-11:10 AM and 3:20-4:20 PM CT
Location: New Orleans Convention Center
