Automatic Data Augmentation for 3D Medical Image Segmentation

Various medical image segmentation models have been created in the past few years for both supervised and semi-supervised tasks. Since variability and diversity of data play a huge role in the accuracy of such models, choosing the most appropriate data augmentation strategy for each organ or tumor is very important in improving the model outcomes. However, most of these data augmentation strategies used in UNet and its variants are simple, custom-made, and not fully optimal. This manual selection of data augmentation strategies limits the possibility for improvement of the accuracy of the medical image segmentation.
As a further improvement on this, Yang et al [1] proposed a technique for the automation data augmentation strategy selection using reinforcement learning. However, this technique requires a high computational time and only tells about the probability of each augmentation strategy. To overcome this issue, Xu et al. [2] proposed an automatic data augmentation strategy framework called ASNG, which searches for the most optimal augmentation strategy by formulating a bi-level optimization algorithm. This framework also designed a search space that includes the fixed magnitude of the operation and the interval of magnitude. Moreover, this work showcases the dynamic change of strategies during training as per requirement, achieving a state-of-the-art performance.
The thesis aims to research the following:
1. How different data augmentation strategies influence the segmentation performance in different data conditions. The data conditions are based on object shape/size, number of objects, mean and standard deviation of the intensity, and so on.
2. Perform analysis of the correlation of the data conditions and the augmentation methods. For this, first, a search space is to be defined which includes typical augmentation strategies with varying hyperparameters, similar to the work done by He et al [3] and Cubuk et al [4]. The performance of different augmentation methods is tested on a segmentation model built using the MONAI framework.
3. Implement an algorithm for searching strategy using machine learning regressor models like support vector regressor that will lead to automatic planning of the augmentation pipeline based on the target dataset. The dataset used for the algorithm development is AMOS [5]. AMOS is a large-scale, diverse, clinical dataset comprising CT and MRI scans for 15 abdominal organ segmentation.