Data Balancing Based on Pre-Training Strategy for Liver Segmentation from CT Scans

Abstract

Data imbalance is often encountered in deep learning process and is harmful to model training. The imbalance of hard and easy samples in training datasets often occurs in the segmentation tasks from Contrast Tomography (CT) scans. However, due to the strong similarity between adjacent slices in volumes and different segmentation tasks (the same slice may be classified as a hard sample in liver segmentation task, but an easy sample in the kidney or spleen segmentation task), it is hard to solve this imbalance of training dataset using traditional methods. In this work, we use a pre-training strategy to distinguish hard and easy samples, and then increase the proportion of hard slices in training dataset, which could mitigate imbalance of hard samples and easy samples in training dataset, and enhance the contribution of hard samples in training process. Our experiments on liver, kidney and spleen segmentation show that increasing the ratio of hard samples in the training dataset could enhance the prediction ability of model by improving its ability to deal with hard samples. The main contribution of this work is the application of pre-training strategy, which enables us to select training samples online according to different tasks and to ease data imbalance in the training dataset.

Subset selection strategy-based pancreas segmentation in CT

“…We also conducted experiments on the NIH dataset using the method in Zhang et al (17). We set the threshold to 0.6783, the minimum dice score of baselines, to distinguish hard and easy samples.…”

“…However, these methods are unable to address the imbalance problem between hard slices and easy slices. Zhang et al used the pretraining method to handle hard-to-easy imbalance in liver segmentation (17), but it required a two-stage training strategy and extra datasets.…”

“…In addition, it is observed that the majority of negative slices are easier to segment than positive ones (Figure 2); therefore, PNSS could also alleviate the hard-to-easy imbalance problem to a certain degree. Inspired by our previous study (17), we propose the HESS Unlike our research group's previous work (17), PNSS and HESS emphasize dynamic adjustment. Especially for HESS, the threshold for distinguishing hard and easy slices between positive slices is automatically changed according to the training state instead of setting a fixed value.…”

“…Following our previous work (17), we used the DSC to assess the difficulty of slices in the training process, with a ( )…”

Special Features on Intelligent Imaging and Analysis

“…It is often noticed that non-balanced datasets lead to a biased estimation model. Zhang et al [23] proposed a pre-training strategy to address this problem encountered in the liver segmentation from computed tomography (CT) scans. Zheng et al…”