Abstract
Data influence methods like TracIn provide critical insights into model behavior by identifying which training examples most significantly impact specific predictions. This paper extends TracIn to the domain of 3D medical image segmentation, specifically targeting brain tumor identification in MRI. By introducing Dice-based gradients and label-specific influence calculations, the authors adapt the method for multi-class segmentation tasks. Experimental results on the BraTS 2019 dataset reveal that while the method is robust to data augmentations, it exhibits low statistical stability across different training runs, likely due to a model navigating a flat loss landscape.
Introduction
Deep learning models are often criticized as "black boxes," a significant hurdle for their adoption in high-stakes clinical environments. While feature attribution methods (e.g., saliency maps) explain where a model focuses, data influence methods explain why a model learned certain behaviors from its training set. This study bridges the gap by applying these methods to semantic segmentation, a task significantly more complex than standard classification.
Methods
The researchers adapted the TracIn algorithm, which estimates influence by tracking loss changes during Stochastic Gradient Descent (SGD). Two primary modifications were implemented for segmentation:
- Dice-based Gradients: Utilizing gradients of the Dice Coefficient instead of cross-entropy to align with standard segmentation performance metrics.
- Label-Specific Influence: Calculating influence scores for individual tissue classes (e.g., tumor core vs. edema) to prevent dominant labels from masking the influence of others.
The study utilized a 2D UNet architecture focused on central brain slices to manage the high computational overhead of influence functions.
Datasets
The BraTS 2019 challenge dataset was used, comprising 259 High-Grade Glioma (HGG) cases. The model was trained on four MRI modalities (T1, T1Gd, T2, T2-FLAIR) to segment three labels: GD-enhancing tumor (ET), peritumoral edema (ED), and necrotic/non-enhancing tumor core (NCR/NET).
Key Results and Insights
The extended TracIn method was evaluated through several consistency and robustness tests:
- Self-Influence: The self-influence matrix exhibited a prominent diagonal, confirming that training examples are typically the strongest proponents for their own predictions.
- Label Separation: Separating influence by label provided a more granular view of data impact compared to a mean Dice approach, preventing tissue types from "washing out" one another.
- Transformation Robustness: The method showed high consistency (cosine similarity > 0.93) when comparing original images to their augmented (rotated or mirrored) counterparts, proving the influence calculation respects learned symmetries.
- Statistical Robustness: A significant finding was the low stability across training runs. For the enhancing tumor label, cosine similarity between influence vectors from different runs was only 0.49 ± 0.23, suggesting that the specific training path (random seed) significantly alters which images are deemed "influential."
Discussion and Conclusion
The study highlights a "flat loss landscape" in segmentation models, where many training examples appear functionally interchangeable. While the method is robust to data transformations, its sensitivity to training runs poses challenges for clinical reliability; if the "most important" training case changes every time a model is retrained, radiologist trust may be undermined. Future work should address the high computational cost of 3D influence functions and focus on measuring the faithfulness of these explanations to the model's underlying decision-making process.
Summary updated