![\"Radiomics](\"https://raw.githubusercontent.com/jpjuvo/RSNA-MICCAI-Brain-Tumor-Classification/main/media/linreg-feat-model.png\")
Did anybody use pyradiomic library for feature extraction? If you did were you able to get a good auc? Did you first use a segmentation model and then used the segmentation masks and original images as input to the pyradiomic library feature extractor?
", "messages": [ { "id": "1551511", "postDate": "10/20/2021 17:24:18", "content": "Did anybody use pyradiomic library for feature extraction? If you did were you able to get a good auc? Did you first use a segmentation model and then used the segmentation masks and original images as input to the pyradiomic library feature extractor?
", "rawMarkdown": "Did anybody use pyradiomic library for feature extraction? If you did were you able to get a good auc? Did you first use a segmentation model and then used the segmentation masks and original images as input to the pyradiomic library feature extractor?", "votes": null }, { "id": "1551635", "postDate": "10/20/2021 19:10:19", "content": "Yes, we tried Pyradiomics features. We trained a 3D tumor segmentation model and extracted Radiomics features from the OOF segmentations.
\nFeatures that seemed to work were:
\nLinear models such as LinReg, Lasso & Ridge seemed pretty stable on all five CV folds (AUC avg ~0.63 std < 0.01). We added random 0.2 * STD-noise augmentations to features while fitting to increase robustness.
\nHowever, our private score was relatively low (AUC 0.55), so this didn't generalize.
@qitvision a really thoroughly documented repo! Thank you for sharing it!
", "rawMarkdown": "qitvision a really thoroughly documented repo! Thank you for sharing it!", "votes": null }, { "id": "1551673", "postDate": "10/20/2021 20:08:31", "content": "Thank you!
", "rawMarkdown": "Thank you!", "votes": null } ], "comments": [ { "id": 1551635, "author_name": "qitvision", "author_url": "", "post_date": "10/20/2021 19:10:19", "content": "Yes, we tried Pyradiomics features. We trained a 3D tumor segmentation model and extracted Radiomics features from the OOF segmentations.
\nFeatures that seemed to work were:
\nLinear models such as LinReg, Lasso & Ridge seemed pretty stable on all five CV folds (AUC avg ~0.63 std < 0.01). We added random 0.2 * STD-noise augmentations to features while fitting to increase robustness.
\nHowever, our private score was relatively low (AUC 0.55), so this didn't generalize.
@qitvision a really thoroughly documented repo! Thank you for sharing it!
", "votes": null, "replies": [] }, { "id": 1551673, "author_name": "qitvision", "author_url": "", "post_date": "10/20/2021 20:08:31", "content": "Thank you!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1551511": "Did anybody use pyradiomic library for feature extraction? If you did were you able to get a good auc? Did you first use a segmentation model and then used the segmentation masks and original images as input to the pyradiomic library feature extractor?", "1551635": "Yes, we tried Pyradiomics features. We trained a 3D tumor segmentation model and extracted Radiomics features from the OOF segmentations.\n\nFeatures that seemed to work were:\n- Tumor location in registered brain coordinates and presence in different parts of the brain (handcrafted features)\n- Tumor 3D shape features such as sphericity and elongation (Pyradiomics features)\n- First-order features such as energy (Pyradiomics features)\n\nLinear models such as LinReg, Lasso & Ridge seemed pretty stable on all five CV folds (AUC avg ~0.63 std < 0.01). We added random 0.2 * STD-noise augmentations to features while fitting to increase robustness.\nHowever, our private score was relatively low (AUC 0.55), so this didn't generalize.\n\n\n\n[code link](https://github.com/jpjuvo/RSNA-MICCAI-Brain-Tumor-Classification)", "1551655": "qitvision a really thoroughly documented repo! Thank you for sharing it!", "1551673": "Thank you!" }, "source": "meta" }