Introduction
In medical image analysis, the performance of deep neural networks for multi-organ segmentation is often limited by the quality of training data and the strategy used to present that data to the model. While significant research focuses on architectural improvements, the selection of training samples remains a critical yet under-explored factor. Standard training typically relies on uniform random sampling, which may overlook difficult cases. Conversely, "hard sample mining" techniques like Online Hard Example Mining (OHEM) prioritize high-loss samples but often fail in medical contexts because the "hardest" samples are frequently those with annotation errors. This paper introduces Relaxed Upper Confident Bound (RUCB), a novel sample selection policy designed to prioritize genuinely difficult anatomy while remaining robust to label noise.
Methodology: From UCB to RUCB
The authors frame the training process as a Multi-Armed Bandit (MAB) problem, where each training slice is treated as an "arm" and the training loss serves as the "reward."
The UCB Baseline
The Upper Confident Bound (UCB) algorithm selects samples by balancing exploitation (choosing samples with high historical loss) and exploration (choosing samples that haven't been seen recently). While effective in stationary environments, standard UCB is "greedy," always selecting the single highest-scoring sample. In the presence of mislabeled data—which produces artificially high, persistent losses—UCB can become trapped, obsessively training on erroneous data.
The RUCB Innovation
To mitigate this, the authors propose Relaxed Upper Confident Bound (RUCB). Instead of selecting the single best arm, RUCB identifies a dynamic range of high-performing arms based on a normal distribution of scores. It then selects a sample randomly from this top-tier group. This stochasticity acts as a shield; by not focusing exclusively on the absolute maximum loss, the model avoids getting stuck on annotation errors and maintains a broader view of the dataset's challenging, but correctly labeled, features.
Experimental Setup
The method was evaluated on a high-quality dataset of 120 abdominal CT scans, targeting the segmentation of 16 distinct structures, including the aorta, pancreas, liver, and various major vessels. A four-fold cross-validation strategy was employed. Despite the dataset's quality, the authors noted that human errors—such as missing small organ parts or mislabeling adjacent vessels—are inevitable and served as the primary test for RUCB's robustness.
Results and Discussion
Quantitative analysis using the Dice-Sørensen Coefficient (DSC) demonstrates that RUCB significantly outperforms both uniform sampling and traditional hard-mining approaches. Notably, OHEM performed worse than random sampling (71.08% vs. 73.55%), confirming that standard hard-mining is detrimental when labels are noisy.
As shown in the table above, RUCB achieved the highest average DSC of 74.75%. The improvements were most pronounced for small, complex organs that are traditionally difficult to segment. For instance, the Adrenal Gland saw a performance jump from 29.33% (Uniform) to 36.76% (RUCB), and the Celiac AA improved from 34.49% to 38.45%. Furthermore, RUCB proved more efficient, reaching higher accuracy levels in fewer iterations than standard UCB.
Conclusion
Relaxed Upper Confident Bound (RUCB) provides a robust framework for training segmentation networks on medical datasets where perfect annotation is unattainable. By introducing a stochastic "relaxation" to the UCB selection policy, the model successfully balances the need to learn from difficult samples without being misled by label noise. This approach is particularly beneficial for small organ segmentation and offers a more efficient training path than traditional uniform or greedy selection methods.
Summary updated