Training medical image segmentation networks on small datasets
We have developed several methods of training neural networks for slice-based segmentation of tissues in 3D CT scans and other medical image modalities. These work especially well for small datasets, for instance, datasets containing only 20 CT scans. We achieve this by using domain-specific pre-processing and augmentation such as using polar coordinates for elliptically distributed tissues, or by embedding depth information to the 2D input images to the network. We found that these approaches generally improve segmentation accuracy and network convergence times for various types of medical images.
Investigating AI fairness in medical image segmentation
We've investigated the fairness of existing methods for skin lesion segmentation using neural networks in dermatological image analysis. Our focus lies in uncovering potential biases within commonly used segmentation models, particularly concerning skin color. Through statistical evaluation techniques, including skin tone estimation methods like Fitzpatrick skin types and Individual Typology Angle estimation, we've established a correlation between segmentation performance and skin color across diverse datasets.