Browsing by Subject "Data augmentation"

The resurgence of Convolutional Neural Networks (CNNs) from the early foundational work is largely attributed to the advent of extensive manually labeled datasets, which has made it possible to train high-capacity models with strong generalization capabilities. However, the annotation cost for these datasets is often prohibitive, and so training CNNs on limited data in a fully-supervised setting remains a crucial problem. Data augmentation is a promising direction for improving generalization in scarce data settings.

We study foundational augmentation techniques, including Mixed Sample Data Augmentations (MSDAs) and a no-parameter variant of RandAugment termed Preset-RandAugment, in the fully supervised scenario. We observe that Preset-RandAugment excels in limited-data contexts while MSDAs are moderately effective. In order to explain this behaviour, we refine ideas about diversity and realism from prior work and propose new ways to measure them. We postulate an additional property when data is limited: augmentations should encourage faster convergence by helping the model learn stable and invariant low-level features, focusing on less class-specific patterns. We explain the effectiveness of Preset-RandAugment in terms of these properties and identify low-level feature transforms as a key contributor to performance.

Building on these insights, we introduce a novel augmentation technique called RandMSAugment that integrates complementary strengths of existing methods. It combines low-level feature transforms from Preset-RandAugment with interpolation and cut-and-paste from MSDA. We improve image diversity through added stochasticity in the mixing process. RandMSAugment significantly outperforms the competition on CIFAR-100, STL-10, and Tiny-Imagenet. With very small training sets (4, 25, 100 samples/class), RandMSAugment achieves compelling performance gains between 4.1\% and 6.75\%. Even with more training data (500 samples/class) we improve performance by 1.03\% to 2.47\%. We also incorporate RandMSAugment augmentations into a semi-supervised learning (SSL) framework and show promising improvements over the state-of-the-art SSL method, FlexMatch. The improvements are more significant when the number of labeled samples is smaller. RandMSAugment does not require hyperparameter tuning, extra validation data, or cumbersome optimizations.

Finally, we combine RandMSAugment with another powerful generalization tool, ensembling, for fully-supervised training with limited samples. We show additonal improvements on the 3 classification benchmarks, which range between 2\% and 5\%. We empirically demonstrate that the gains due to ensembling are larger when the individual networks have moderate accuracies \ie outside of the low and high extremes.Furthermore, we introduce a simulation tool capable of providing insights about the maximum accuracy achievable through ensembling, under various conditions.

IDH is a gene that heavily affects the treatment response of gliomas and is associated with patient’s prognosis. Convolutional neural networks (CNNs) showed great potential in predicting IDH mutations. However, these CNN models require time-intensive image-preprocessing before being used for predictions. There are two main purpose of this study. The first purpose is to investigate the feasibility of applying a novel convolutional neural network based on the Inception-ResNet to reduce image preprocessing steps and improve accuracy in IDH mutation prediction. The second purpose is to evaluate different data augmentation methods on brain studies.

MR images of 103 patients were selected from The Cancer Imaging Archive (TCIA). Intensity normalization of every individual slice is the only image pre-processing step. The T1w post-contrast, FLAIR, and T2w images at the same slice location were grouped together and considered as one training sample. This give rise to 209 IDH-mutant samples from 42 patients and 356 IDH-wild-type samples from 61 patients that were randomly selected to become training, validation, and test sets. To avoid overfitting in the model performance, data augmentation methods were applied individually to both training and validation sets in each training. The augmentation methods included duplication, noise addition, rotation, translation, cropping and mirroring. Images from one sample were fed to different input channels of Inception-ResNet, and the predictions were based on the extracted features and the patient’s age at diagnosis. Prediction accuracy was used to assess the performance of different augmentation methods.

With only intensity normalization, the proposed model using training sets augmented by rotation and noise addition achieved the IDH prediction accuracies of 91.8% and 91.8%, respectively. On the same training, validation and test sets, the proposed model trained on data augmented by duplication, cropping, translation, and mirroring gave accuracies of 81.6%, 79.6%, 83.7%, and 85.7%, respectively.

This work investigated the feasibility of the application of the novel convolutional neural network based on the Inception-ResNet on IDH mutation prediction, and high accuracies can be achieved with only intensity normalization as image preprocessing. Among all data augmentation methods, noise addition and rotation shows better performance and might suggest potential value for other clinical applications using machine learning algorithms.