{ "id": 475964, "title": "7th solution", "url": "/competitions/blood-vessel-segmentation/discussion/475964", "author_name": "sakaku", "post_date": "2024-02-10T15:51:31.473000", "votes": 16, "comment_count": 0, "views": 0, "content": "

First and foremost, I'd like to extend my gratitude to Kaggle and the competition organizers for creating such a compelling event. Despite joining the contest relatively late, I was able to quickly get up to speed thanks to insightful discussions by @hengck23 , the informative videos by @yoyobar , and the vibrant exchanges within the community.

\n

Overview

\n

I employed a hybrid approach for the model, utilizing 2.5D images as input. The architecture combines a 2D Unet framework with 3D convolutional layers. From my research and the community's insights, it seemed that a full 3D model might offer superior results compared to 2D models. However, due to the high computational costs associated with 3D models, I choose a blend of 3D convolutions within a 2D Unet structure, striking a balance between efficiency and performance.

\n

Data Preparation

\n

Generating 3D Rotational Slices

\n

I augmented the dataset with 3D rotation. The process begins by assembling the images into a 3D volume, followed by rotating two axes and extracting slices along the remaining axis. The rotation angles used are as follows:

\n
rotation_angles = [\n    [, ], [, -], [-, ], [-, -],\n    [, ], [, -], [-, ], [-, -],\n    [, ], [, -], [-, ], [-, -]\n]\n
\n

Post-rotation, some slices exhibited increased areas of black background. To maintain data quality, I retained only those slices where the target segmentation was present and the black background constituted less than 50% of the slice area.

\n

Rotate data sample

\n

\n

Pseudo Labeling

\n

The pseudo labeling process involved:

\n
    \n
  1. Generating additional slices for kidney1 and kidney3 using the aforementioned technique.
    2. \n
  2. Training a model with the augmented dataset.
    3. \n
  3. Applying the model to pseudo label kidney2, followed by generating extra slices for it in a similar manner.
    4. \n
\n

Model

\n

Architecture

\n

\n

Training and Inference Details

\n

Training

\n

The training setup was as follows:

\n
    \n
  1. The model takes a 3-channel 2.5D image as input and outputs a 3-channel prediction.
    2. \n
  2. Normalize the input base on the std of each kidney
    3. \n
  3. I used a combination of loss functions: BCEWithLogitsLoss, DiceLoss, and FocalLoss. The loss for each of the three channels was calculated separately, with the middle channel assigned a higher weight of 0.9.
    4. \n
  4. Optimizer: AdamW
    5. \n
  5. Scheduler: CosineAnnealingLR
    6. \n
  6. Images were cropped to a size of 1024x1024 for processing.
    7. \n
\n

Inference

\n

For inference:

\n
    \n
  1. A single model was used for predictions.
    2. \n
  2. The model operated at the original image resolution.
    3. \n
  3. Similar to training, the input comprised 3-channel 2.5D images, with the output being a 3-channel prediction, primarily focusing on the middle channel.
    4. \n
  4. Normalize the input base on the std of each kidney
    5. \n
  5. Predictions were made along the x, y, and z axes, and the results were averaged.
    6. \n
  6. Test Time Augmentation (TTA) included horizontal flipping.
    7. \n
  7. Post-processing steps involved applying thresholds of 0.2, followed by a 3D closing operation.
    8. \n
", "messages": [ { "id": 2646004, "postDate": "2024-02-10T15:51:31.473Z", "content": "

First and foremost, I'd like to extend my gratitude to Kaggle and the competition organizers for creating such a compelling event. Despite joining the contest relatively late, I was able to quickly get up to speed thanks to insightful discussions by @hengck23 , the informative videos by @yoyobar , and the vibrant exchanges within the community.

\n

Overview

\n

I employed a hybrid approach for the model, utilizing 2.5D images as input. The architecture combines a 2D Unet framework with 3D convolutional layers. From my research and the community's insights, it seemed that a full 3D model might offer superior results compared to 2D models. However, due to the high computational costs associated with 3D models, I choose a blend of 3D convolutions within a 2D Unet structure, striking a balance between efficiency and performance.

\n

Data Preparation

\n

Generating 3D Rotational Slices

\n

I augmented the dataset with 3D rotation. The process begins by assembling the images into a 3D volume, followed by rotating two axes and extracting slices along the remaining axis. The rotation angles used are as follows:

\n
rotation_angles = [\n    [, ], [, -], [-, ], [-, -],\n    [, ], [, -], [-, ], [-, -],\n    [, ], [, -], [-, ], [-, -]\n]\n
\n

Post-rotation, some slices exhibited increased areas of black background. To maintain data quality, I retained only those slices where the target segmentation was present and the black background constituted less than 50% of the slice area.

\n

Rotate data sample

\n

\n

Pseudo Labeling

\n

The pseudo labeling process involved:

\n
    \n
  1. Generating additional slices for kidney1 and kidney3 using the aforementioned technique.
    2. \n
  2. Training a model with the augmented dataset.
    3. \n
  3. Applying the model to pseudo label kidney2, followed by generating extra slices for it in a similar manner.
    4. \n
\n

Model

\n

Architecture

\n

\n

Training and Inference Details

\n

Training

\n

The training setup was as follows:

\n
    \n
  1. The model takes a 3-channel 2.5D image as input and outputs a 3-channel prediction.
    2. \n
  2. Normalize the input base on the std of each kidney
    3. \n
  3. I used a combination of loss functions: BCEWithLogitsLoss, DiceLoss, and FocalLoss. The loss for each of the three channels was calculated separately, with the middle channel assigned a higher weight of 0.9.
    4. \n
  4. Optimizer: AdamW
    5. \n
  5. Scheduler: CosineAnnealingLR
    6. \n
  6. Images were cropped to a size of 1024x1024 for processing.
    7. \n
\n

Inference

\n

For inference:

\n
    \n
  1. A single model was used for predictions.
    2. \n
  2. The model operated at the original image resolution.
    3. \n
  3. Similar to training, the input comprised 3-channel 2.5D images, with the output being a 3-channel prediction, primarily focusing on the middle channel.
    4. \n
  4. Normalize the input base on the std of each kidney
    5. \n
  5. Predictions were made along the x, y, and z axes, and the results were averaged.
    6. \n
  6. Test Time Augmentation (TTA) included horizontal flipping.
    7. \n
  7. Post-processing steps involved applying thresholds of 0.2, followed by a 3D closing operation.
    8. \n
", "rawMarkdown": "First and foremost, I'd like to extend my gratitude to Kaggle and the competition organizers for creating such a compelling event. Despite joining the contest relatively late, I was able to quickly get up to speed thanks to insightful discussions by @hengck23 , the informative videos by @yoyobar , and the vibrant exchanges within the community.\n\n## Overview\n\nI employed a hybrid approach for the model, utilizing 2.5D images as input. The architecture combines a 2D Unet framework with 3D convolutional layers. From my research and the community's insights, it seemed that a full 3D model might offer superior results compared to 2D models. However, due to the high computational costs associated with 3D models, I choose a blend of 3D convolutions within a 2D Unet structure, striking a balance between efficiency and performance.\n\n## Data Preparation\n\n### Generating 3D Rotational Slices\n\nI augmented the dataset with 3D rotation. The process begins by assembling the images into a 3D volume, followed by rotating two axes and extracting slices along the remaining axis. The rotation angles used are as follows:\n\n```python\nrotation_angles = [\n [10, 10], [10, -10], [-10, 10], [-10, -10],\n [30, 30], [30, -30], [-30, 30], [-30, -30],\n [45, 45], [45, -45], [-45, 45], [-45, -45]\n]\n```\n\nPost-rotation, some slices exhibited increased areas of black background. To maintain data quality, I retained only those slices where the target segmentation was present and the black background constituted less than 50% of the slice area. \n#### Rotate data sample\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2369671%2F2d8dc94653829769d2cc750903ab67de%2Fk1_z_pseudo_rotset7_1257.jpg?generation=1707579928485450&alt=media)\n### Pseudo Labeling\n\nThe pseudo labeling process involved:\n1. Generating additional slices for kidney1 and kidney3 using the aforementioned technique.\n2. Training a model with the augmented dataset.\n3. Applying the model to pseudo label kidney2, followed by generating extra slices for it in a similar manner.\n\n## Model\n\n### Architecture\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2369671%2F2152324997361e303fb5fdfed3811b7d%2FScreenshot%202024-02-10%20at%2023.52.55.png?generation=1707579666652474&alt=media)\n### Training and Inference Details\n\n#### Training\n\nThe training setup was as follows:\n1. The model takes a 3-channel 2.5D image as input and outputs a 3-channel prediction.\n2. Normalize the input base on the std of each kidney\n2. I used a combination of loss functions: BCEWithLogitsLoss, DiceLoss, and FocalLoss. The loss for each of the three channels was calculated separately, with the middle channel assigned a higher weight of 0.9.\n3. Optimizer: AdamW\n4. Scheduler: CosineAnnealingLR\n5. Images were cropped to a size of 1024x1024 for processing.\n\n#### Inference\n\nFor inference:\n1. A single model was used for predictions.\n2. The model operated at the original image resolution.\n3. Similar to training, the input comprised 3-channel 2.5D images, with the output being a 3-channel prediction, primarily focusing on the middle channel.\n2. Normalize the input base on the std of each kidney\n4. Predictions were made along the x, y, and z axes, and the results were averaged.\n5. Test Time Augmentation (TTA) included horizontal flipping.\n6. Post-processing steps involved applying thresholds of 0.2, followed by a 3D closing operation.\n", "votes": 16 } ], "comments": [], "raw_markdown_by_id": { "2646004": "First and foremost, I'd like to extend my gratitude to Kaggle and the competition organizers for creating such a compelling event. Despite joining the contest relatively late, I was able to quickly get up to speed thanks to insightful discussions by @hengck23 , the informative videos by @yoyobar , and the vibrant exchanges within the community.\n\n## Overview\n\nI employed a hybrid approach for the model, utilizing 2.5D images as input. The architecture combines a 2D Unet framework with 3D convolutional layers. From my research and the community's insights, it seemed that a full 3D model might offer superior results compared to 2D models. However, due to the high computational costs associated with 3D models, I choose a blend of 3D convolutions within a 2D Unet structure, striking a balance between efficiency and performance.\n\n## Data Preparation\n\n### Generating 3D Rotational Slices\n\nI augmented the dataset with 3D rotation. The process begins by assembling the images into a 3D volume, followed by rotating two axes and extracting slices along the remaining axis. The rotation angles used are as follows:\n\n```python\nrotation_angles = [\n [10, 10], [10, -10], [-10, 10], [-10, -10],\n [30, 30], [30, -30], [-30, 30], [-30, -30],\n [45, 45], [45, -45], [-45, 45], [-45, -45]\n]\n```\n\nPost-rotation, some slices exhibited increased areas of black background. To maintain data quality, I retained only those slices where the target segmentation was present and the black background constituted less than 50% of the slice area. \n#### Rotate data sample\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2369671%2F2d8dc94653829769d2cc750903ab67de%2Fk1_z_pseudo_rotset7_1257.jpg?generation=1707579928485450&alt=media)\n### Pseudo Labeling\n\nThe pseudo labeling process involved:\n1. Generating additional slices for kidney1 and kidney3 using the aforementioned technique.\n2. Training a model with the augmented dataset.\n3. Applying the model to pseudo label kidney2, followed by generating extra slices for it in a similar manner.\n\n## Model\n\n### Architecture\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2369671%2F2152324997361e303fb5fdfed3811b7d%2FScreenshot%202024-02-10%20at%2023.52.55.png?generation=1707579666652474&alt=media)\n### Training and Inference Details\n\n#### Training\n\nThe training setup was as follows:\n1. The model takes a 3-channel 2.5D image as input and outputs a 3-channel prediction.\n2. Normalize the input base on the std of each kidney\n2. I used a combination of loss functions: BCEWithLogitsLoss, DiceLoss, and FocalLoss. The loss for each of the three channels was calculated separately, with the middle channel assigned a higher weight of 0.9.\n3. Optimizer: AdamW\n4. Scheduler: CosineAnnealingLR\n5. Images were cropped to a size of 1024x1024 for processing.\n\n#### Inference\n\nFor inference:\n1. A single model was used for predictions.\n2. The model operated at the original image resolution.\n3. Similar to training, the input comprised 3-channel 2.5D images, with the output being a 3-channel prediction, primarily focusing on the middle channel.\n2. Normalize the input base on the std of each kidney\n4. Predictions were made along the x, y, and z axes, and the results were averaged.\n5. Test Time Augmentation (TTA) included horizontal flipping.\n6. Post-processing steps involved applying thresholds of 0.2, followed by a 3D closing operation.\n" } }