What's Behind Auto-Segmentation Models: An Interpretability Analysis of Calculation Logic of Segmentation Model for Brain Tumor
H Chen1*, D Ban2, X Qi3, (1) Xi'an University of Posts and telecommunications, Xi'an, 61, CN, (2) Xi'an University of Posts and telecommunications, Xi'an, 61, CN, (3) UCLA School of Medicine, Los Angeles, CA
Presentations
PO-GePV-M-38 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: We proposed a workflow to exam the implicit calculation logic behind segmentation models (U-NET，XGBoost) to better understand the underlying algorithms. We attempted to fit two segmentation models respectively through decision trees in a predefined feature space. The highly similar decision paths generated from these two models will be extracted to discover the associations among features.
Methods: 369 patients from the BraTS20 were used and regrouped into 3 subgroups: long-term survival (>15 months), mid-term survival (10-15 months), and short-term survival (<10 months). Each patient had four preoperative multimodal MR scans: T1, T2, Flair, T1ce. A total of 444 radiomics features (74 first order features, 224 features from Gray Level Cooccurence Matrix and 146 features from Gray Level Run Length Matrix) on MR images of 4 sequences were extracted from 295 patients to train an ensemble learning segmentation model and a U-NET segmentation model respectively. The decision tree was used to fit above models. The decision paths for segmenting each subregion of glioma, including the non-enhancing tumor region (NET) and the edema area (ED) respectively for the long-term subgroup, were filtered out to extract the most common feature associations.
Results: A total of 14,636 decision paths in the segmentation models between NET and ED (6,836 for NET and 7,800 for ED) were examed. There were three highly similar decision paths between two models, which include 8 features . The most popular feature associations between NET and ED were identified respectively, in-cluding AutoCorrelation_180_f, AutoCorrelation_90_f and median_t1c, and AutoCorrelation_180_f, Con-trast_180_t1c, AutoCorrelation_270_t1c, min_t2, etc.
Conclusion: We demonstrated a novel workflow to analyze the underlying workflow of feature based auto-segmentation models. This method allows us to interpret the decision logic of the black box model in the boundary seg-menting process, and show a promising way to explore and improve the interpretability of machine learning model.
Funding Support, Disclosures, and Conflict of Interest: The Natural Science Basic Research Program of Shaanxi Province of China [2019JM-365]; The Scientific Research Program Funded by Shaanxi Provincial Education Department of China[17JK0701]
ePosters
Keywords
MRI, Segmentation, Quantitative Imaging
Taxonomy
IM/TH- Image Segmentation Techniques: Machine Learning
Contact Email