Analyzing Domain Shift When Using Additional Data for the MICCAI KiTS23 Challenge

Using additional training data is known to improve the results, especially for medical image 3D segmentation where there is a lack of training material and the model needs to generalize well from few available data. Unlike transfer learning in which a model pretrained on huge datasets is fine-tuned for a specific task using limited data, we research the case in which we acquire supplementary training material and combine it with the original training data. However, the new data could have been obtained using other instruments and preprocessed such its distribution is significantly different from the target domain. Therefore, we study techniques which ameliorate domain shift during training so that the additional data becomes better usable for preprocessing and training together with the original data. We opt for using statistical criteria for reducing the distribution shift for domain adaptation in the context of having more data from the target domain than additional training data. Our results show that transforming the additional data using histogram matching has better results than using simple normalization. We achieved the \(5^{th}\) place on the official test dataset with a Dice score of 0.807 and Surface Dice of 0.691. On the validation set, we additionally report the Dice score for cysts (0.512) and kidney (0.946) besides the official metrics.