Deep Learning Training Strategies for Severely Imbalanced Data in Organ Segmentation Tasks

Abstract

Radiotherapy is one of the common methods for cancer treatment. Developing a radiotherapy plan requires professional medical physicists or physicians to manually contour the organ boundaries in CT series, which is time-and labor-consuming. If artificial intelligence (AI) could assist with the task, it could alleviate the workload of medical staff, especially when medical resources are tight. We propose an AI-based automatic organ segmentation system trained by clinical datasets. However, this task is prone to be non-robust models in CT image series where the background occupies the majority of the scene. To remedy such data imbalance situation, we propose adopting three strategies during the model training steps: region classification, knowledge discovery in database, and sampler. The major segmentation task is based on U-Net and ResNet34 model where all convolution layers and batch normalization are replaced with group normalization and weight standardization to ensure effectiveness in small-batch data training. In this study, 33 organs throughout the body were segmented. The ablation experiments were conducted to prove all the training models have better performance than the original method. In the future, if a hospital needs to train model with their own private datasets, the three above strategies can be adopted to prevent unsuccessful training.