{ "id": 359128, "title": "Recurrent neural networks trained on Effnetv2 embeddings", "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/359128", "author_name": "_lev_lipinski", "post_date": "2022-10-10T20:36:35.801000", "votes": 9, "comment_count": 0, "views": 0, "content": "

Hi all,

\n

I've explored using recurrent neural networks for this challege, you can find my notebook here: Two-stage LSTM-GRU on EffNetv2 embeddings\n. I trained an LSTM to make slicewise vertebrae and fracture predictions using Effnetv2 embeddings from [train] PyTorch-EffNetV2 baseline CV:0.49. The intuition is that a sequential model can have the additional capability to capture information from other slices of the patient, instead of simply taking into account a single slice. I trained a GRU to make the final predictions using the 14-dim (7-dim vertebrae, 7-dim fracture predictions) sequences (i.e. slices) for each patient. You can find my approach below.

\n

Although the models seem to work, it's weird that I could not improve the performance of the notebook [train] PyTorch-EffNetV2 baseline CV:0.49, I basically have similar results. Any suggestions would be appreciated on where I could improve the results.

\n

\"\"

", "messages": [ { "id": 1981526, "postDate": "2022-10-10T20:36:35.803Z", "content": "

Hi all,

\n

I've explored using recurrent neural networks for this challege, you can find my notebook here: Two-stage LSTM-GRU on EffNetv2 embeddings\n. I trained an LSTM to make slicewise vertebrae and fracture predictions using Effnetv2 embeddings from [train] PyTorch-EffNetV2 baseline CV:0.49. The intuition is that a sequential model can have the additional capability to capture information from other slices of the patient, instead of simply taking into account a single slice. I trained a GRU to make the final predictions using the 14-dim (7-dim vertebrae, 7-dim fracture predictions) sequences (i.e. slices) for each patient. You can find my approach below.

\n

Although the models seem to work, it's weird that I could not improve the performance of the notebook [train] PyTorch-EffNetV2 baseline CV:0.49, I basically have similar results. Any suggestions would be appreciated on where I could improve the results.

\n

\"\"

", "rawMarkdown": "Hi all,\n\nI've explored using recurrent neural networks for this challege, you can find my notebook here: [Two-stage LSTM-GRU on EffNetv2 embeddings\n](https://www.kaggle.com/code/leventelippenszky/two-stage-lstm-gru-on-effnetv2-embeddings). I trained an LSTM to make slicewise vertebrae and fracture predictions using Effnetv2 embeddings from [[train] PyTorch-EffNetV2 baseline CV:0.49](https://www.kaggle.com/code/vslaykovsky/train-pytorch-effnetv2-baseline-cv-0-49). The intuition is that a sequential model can have the additional capability to capture information from other slices of the patient, instead of simply taking into account a single slice. I trained a GRU to make the final predictions using the 14-dim (7-dim vertebrae, 7-dim fracture predictions) sequences (i.e. slices) for each patient. You can find my approach below.\n\nAlthough the models seem to work, it's weird that I could not improve the performance of the notebook [[train] PyTorch-EffNetV2 baseline CV:0.49](https://www.kaggle.com/code/vslaykovsky/train-pytorch-effnetv2-baseline-cv-0-49), I basically have similar results. Any suggestions would be appreciated on where I could improve the results.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3082575%2F57c0dd16c3d13919ff04fb847262fc0b%2Flstm_gru_rsna_2022.PNG?generation=1665433864782315&alt=media)", "votes": 9 } ], "comments": [], "raw_markdown_by_id": { "1981526": "Hi all,\n\nI've explored using recurrent neural networks for this challege, you can find my notebook here: [Two-stage LSTM-GRU on EffNetv2 embeddings\n](https://www.kaggle.com/code/leventelippenszky/two-stage-lstm-gru-on-effnetv2-embeddings). I trained an LSTM to make slicewise vertebrae and fracture predictions using Effnetv2 embeddings from [[train] PyTorch-EffNetV2 baseline CV:0.49](https://www.kaggle.com/code/vslaykovsky/train-pytorch-effnetv2-baseline-cv-0-49). The intuition is that a sequential model can have the additional capability to capture information from other slices of the patient, instead of simply taking into account a single slice. I trained a GRU to make the final predictions using the 14-dim (7-dim vertebrae, 7-dim fracture predictions) sequences (i.e. slices) for each patient. You can find my approach below.\n\nAlthough the models seem to work, it's weird that I could not improve the performance of the notebook [[train] PyTorch-EffNetV2 baseline CV:0.49](https://www.kaggle.com/code/vslaykovsky/train-pytorch-effnetv2-baseline-cv-0-49), I basically have similar results. Any suggestions would be appreciated on where I could improve the results.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3082575%2F57c0dd16c3d13919ff04fb847262fc0b%2Flstm_gru_rsna_2022.PNG?generation=1665433864782315&alt=media)" } }