Baseline

Motivated by the well-known learning without forgetting [1], we develop an embarrassingly simple but effective continual learning method as the baseline, which contains the following four steps

• Step 1. Individually training a liver segmentation nnU-Net [2] model based on the MSD Pancreas Ts (139) dataset.
• Step 2. Using the trained liver segmentation model to infer KITS (210) and obtaining pseudo liver labels. Thus, each case in the KITS (210) has both liver and kidney labels. Then, we use the new labels to train a nnU-Net model that can segment both liver and kidney.
• Step 3. Using the trained model in Step 2 to infer Spleen (41) and obtaining both liver and kidney pseudo labels. Thus, each case in the Spleen (210) has liver, kidney, and spleen labels. Then, we use the new labels to train a nnU-Net model that can segment the three organs.
• Step 4. Using the trained model in Step 3 to infer MSD Pancreas (281) and obtaining both liver, kidney and spleen pseudo labels. Thus, each case in the MSD Pancreas (281) has liver, kidney, spleen, and pancreas labels. Finally, we can obtain the final multi-organ segmentation model by training a nnU-Net with the new labels.

Code and trained models are publicly available here.

Reference

[1] Z. Li and D. Hoiem, “Learning without forgetting,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 40, no. 12, pp. 2935–2947, 2017.

[2] F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein, "nnU-Net: a self-confifiguring method for deep learning" based biomedical image segmentation,” Nature Methods, vol. 18, no. 2, pp. 203–211, 2021.