Worst case prediction: Lung (Average DSC: 0.1925)
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6261540%2F5a67f6179ac3e0ae91b57c3682a268ec%2Flung_3_wsi_worst.png?generation=1663951851101765&alt=media\")
Best case prediction: Prostate (Average DSC: 0.8017)
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6261540%2Fe188e36765bf2c76404a52eac0e9ade0%2Fprostate_0_wsi_best.png?generation=1663951873035576&alt=media\")
First of all, I would like to thank the organizers and Kaggle team for hosting this competition, and congratulate to all the winners! I also want to thank @thedevastator, @hengck23, @jamesphoward, @gunesevitan, @kingjohnson and many other Kagglers for their discussions and contributions. I've been inspired a lot by their insightful discussions and valuable contributions to this competition.
\nI've seen many discussions related to training on tiles vs non-tiles and how it affects the public LB. So, I was curious and decided to do some experiments based on these data settings. I used the UNeXt50 baseline provided by @thedevastator for all my experiments.
\nHere is my methodology in detail: HuBMAP_HPA methodology
\nData:
\nData augmentation: heavy data augmentation pipeline including morphology and color changes
\nModel architecture: UNeXt50 baseline
\nTraining setup:
\nAll trainings were done using GPU quota from Kaggle.
\nResults:
\nWorst case prediction: Lung (Average DSC: 0.1925)
\n
Best case prediction: Prostate (Average DSC: 0.8017)
\n
Inference:
\nMy public and private LB :
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6261540%2F686924ea25aae33c0840f7fc678e24ec%2FScreen%20Shot%202022-09-23%20at%2012.58.19%20PM.png?generation=1663952329198584&alt=media\")
Other things I tried out but failed:
\nConclusion:
\nWith some tweaks, I boosted the UNeXt50 baseline from 0.51322 to 0.73040 in private LB. Training on non-tiles performed better in this case.
Thanks for reading!
", "messages": [ { "id": 1952440, "postDate": "2022-09-23T17:04:26.863Z", "content": "First of all, I would like to thank the organizers and Kaggle team for hosting this competition, and congratulate to all the winners! I also want to thank @thedevastator, @hengck23, @jamesphoward, @gunesevitan, @kingjohnson and many other Kagglers for their discussions and contributions. I've been inspired a lot by their insightful discussions and valuable contributions to this competition.
\nI've seen many discussions related to training on tiles vs non-tiles and how it affects the public LB. So, I was curious and decided to do some experiments based on these data settings. I used the UNeXt50 baseline provided by @thedevastator for all my experiments.
\nHere is my methodology in detail: HuBMAP_HPA methodology
\nData:
\nData augmentation: heavy data augmentation pipeline including morphology and color changes
\nModel architecture: UNeXt50 baseline
\nTraining setup:
\nAll trainings were done using GPU quota from Kaggle.
\nResults:
\nWorst case prediction: Lung (Average DSC: 0.1925)
\n
Best case prediction: Prostate (Average DSC: 0.8017)
\n
Inference:
\nMy public and private LB :
\n
Other things I tried out but failed:
\nConclusion:
\nWith some tweaks, I boosted the UNeXt50 baseline from 0.51322 to 0.73040 in private LB. Training on non-tiles performed better in this case.
Thanks for reading!
", "rawMarkdown": "First of all, I would like to thank the organizers and Kaggle team for hosting this competition, and congratulate to all the winners! I also want to thank @thedevastator, @hengck23, @jamesphoward, @gunesevitan, @kingjohnson and many other Kagglers for their discussions and contributions. I've been inspired a lot by their insightful discussions and valuable contributions to this competition.\n\n---\n\nI've seen many discussions related to training on tiles vs non-tiles and how it affects the public LB. So, I was curious and decided to do some experiments based on these data settings. I used the [UNeXt50 baseline](https://www.kaggle.com/code/thedevastator/training-fastai-baseline) provided by @thedevastator for all my experiments. \n\nHere is my methodology in detail: [HuBMAP_HPA methodology](https://www.kaggle.com/code/nghihuynh/hubmap-hpa-methodology-submission)\n\n---\n\n**Data**: \n* Prepare tile dataset using [Otsu and Triangle thresholding](https://www.kaggle.com/code/nghihuynh/wsi-preprocessing-tiling-tissue-segmentation) to remove a large portion of background vs non-FTUs tissue.\n* Prepare [non-tile dataset](https://www.kaggle.com/code/nghihuynh/wsi-hubmap-dataset-creation-256x256) by cropping the tissue area and rescaling to 256x256, and 512x512\n\n**Data augmentation**: heavy data augmentation pipeline including morphology and color changes\n\n**Model architecture**: UNeXt50 baseline\n\n**Training setup**: \n* Cross-validation: Stratified K-Fold to select training images based on organs. (4 folds)\n* Loss function: asymmetric Lovasz Hinge with the weight of 0.55 for positive prediction, and 0.45 for negative prediction.\n* Number of epochs: at least 20 epochs, at most 80 epochs. Most models converged after 60 epochs.\n\n*All trainings were done using GPU quota from Kaggle.*\n\n**Results**:\n\n* Worst case prediction: Lung (Average DSC: 0.1925)\n\n\n* Best case prediction: Prostate (Average DSC: 0.8017)\n\n\n**Inference**:\n* Tiles: 0.64 (averaging 4 folds)\n* Non-tiles: 0.74 (averaging 4 folds)-for experiments only. I used more ensemble models for submitting.\n\nMy public and private LB : \n\n\n---\n\nOther things I tried out but failed:\n\n* Multi-class segmentation: I'm still not quite sure why it gave poorer results, even though the dice soft was much higher than in binary segmentation models. Any insights are really appreciated!\n* Fine-tuning on synthetic dataset using Gaussian-Laplacian blending: overfitting might be an issue. \n\n---\n**Conclusion**:\nWith some tweaks, I boosted the UNeXt50 baseline from 0.51322 to 0.73040 in private LB. Training on non-tiles performed better in this case. \n\nThanks for reading! ", "votes": 10 } ], "comments": [], "raw_markdown_by_id": { "1952440": "First of all, I would like to thank the organizers and Kaggle team for hosting this competition, and congratulate to all the winners! I also want to thank @thedevastator, @hengck23, @jamesphoward, @gunesevitan, @kingjohnson and many other Kagglers for their discussions and contributions. I've been inspired a lot by their insightful discussions and valuable contributions to this competition.\n\n---\n\nI've seen many discussions related to training on tiles vs non-tiles and how it affects the public LB. So, I was curious and decided to do some experiments based on these data settings. I used the [UNeXt50 baseline](https://www.kaggle.com/code/thedevastator/training-fastai-baseline) provided by @thedevastator for all my experiments. \n\nHere is my methodology in detail: [HuBMAP_HPA methodology](https://www.kaggle.com/code/nghihuynh/hubmap-hpa-methodology-submission)\n\n---\n\n**Data**: \n* Prepare tile dataset using [Otsu and Triangle thresholding](https://www.kaggle.com/code/nghihuynh/wsi-preprocessing-tiling-tissue-segmentation) to remove a large portion of background vs non-FTUs tissue.\n* Prepare [non-tile dataset](https://www.kaggle.com/code/nghihuynh/wsi-hubmap-dataset-creation-256x256) by cropping the tissue area and rescaling to 256x256, and 512x512\n\n**Data augmentation**: heavy data augmentation pipeline including morphology and color changes\n\n**Model architecture**: UNeXt50 baseline\n\n**Training setup**: \n* Cross-validation: Stratified K-Fold to select training images based on organs. (4 folds)\n* Loss function: asymmetric Lovasz Hinge with the weight of 0.55 for positive prediction, and 0.45 for negative prediction.\n* Number of epochs: at least 20 epochs, at most 80 epochs. Most models converged after 60 epochs.\n\n*All trainings were done using GPU quota from Kaggle.*\n\n**Results**:\n\n* Worst case prediction: Lung (Average DSC: 0.1925)\n\n\n* Best case prediction: Prostate (Average DSC: 0.8017)\n\n\n**Inference**:\n* Tiles: 0.64 (averaging 4 folds)\n* Non-tiles: 0.74 (averaging 4 folds)-for experiments only. I used more ensemble models for submitting.\n\nMy public and private LB : \n\n\n---\n\nOther things I tried out but failed:\n\n* Multi-class segmentation: I'm still not quite sure why it gave poorer results, even though the dice soft was much higher than in binary segmentation models. Any insights are really appreciated!\n* Fine-tuning on synthetic dataset using Gaussian-Laplacian blending: overfitting might be an issue. \n\n---\n**Conclusion**:\nWith some tweaks, I boosted the UNeXt50 baseline from 0.51322 to 0.73040 in private LB. Training on non-tiles performed better in this case. \n\nThanks for reading! " } }