Contrastive Learning versus Deformable Data Augmentation in Semantic Organ Segmentation
Hossein Arabi, Habib Zaidi · 2022
Extraction of the underlying patterns/features from the input data is essential for the development of an accurate and robust segmentation model. In this regard, self- or contrastive learning techniques are employed to enhance the performance of deep learning models particularly when the size of the training dataset is suboptimal. This work set out to assess the impact of contrastive learning on the performance of the semantic segmentation model. Moreover, a deformable augmentation approach is introduced to be compared with the contrastive-learning approach. The comparison of the contrastive learning and the proposed deformable data-augmentation technique was conducted on two datasets concerning hippocampus and kidney segmentation from the MR and CT images (Decathlon dataset), respectively. The accuracy of the segmentation for the two datasets was investigated with and without applying contrastive learning and the proposed deformable data-augmentation technique. The deformable data-augmentation technique resulted in Dice indices of 0.920±0.022 for kidney segmentation and 0.898±0.027 for hippocampus segmentation. The contrastive-learning technique exhibited superior performance to the initial model in both datasets with statistically significant differences (p-value <0.03). However, the proposed deformable data-augmentation approach outperformed contrastive learning due to the increased dataset within training. The contrastive learning and the original model exhibited Dice indices of 0.871±0.039 and 0.868±0.042 for kidney segmentation, and 0.872±0.045 and 0.865±0.048 for hippocampus segmentation. The proposed deformable data-augmentation technique dramatically improved the performance of the segmentation model, and a combination of the contrastive-learning and data augmentation would lead to a robust segmentation model with no outliers.