First of all, I would like to express my sincere gratitude to the competition host and the Kaggle staff for organizing such a fascinating competition. I thoroughly enjoyed this competition and learned a great deal in the process!
\nFurthermore, I'd like to thank @hengck23 and @davidlist . @hengck23 's discussion and notebook were the starting point for my solution, and Many of the discussions and comments from @davidlist were extremely helpful.
\nMy solution is simple and straightforward. I performed segmentation using a 3D ConvNeXt-like model and further employed an ensemble of as many models as possible. Subsequently, I computed the centroids of the particles using cc3d and curated the predicted centroids with DBSCAN.
\nI used ground truth masks with an adjusted radius for each particle. As a result, the public leaderboard score jumped 0.02~0.04. Concletely, I adjust mask size as shown in the below table;
\n| apo-ferritin(r=60) | \nbeta-galactosidase(r=90) | \nribosome(r=150) | \nthyroglobulin(r=130) | \nvirus-like-particle(r=135) | \n
|---|---|---|---|---|
| 60/2 | \n90/2 | \n150/3 | \n130/3 | \n135/3 | \n
Since the competition metric requires that the predicted centroid falls within a radius of r×0.5 from the ground truth centroid, I believe that multiplying the radius of each particle by a factor of 0.5 or less is reasonable.
\nIn this section, I explain my model architecture both encoder and decoder.
\nTo implement the encoder, I started with the ConvNeXt, since ConvNeXt is very powerful and fast. Then, I customized the model to suit the task that is to predict small particles. These changes from the original ConvNeXt are below;
\nThe flow of the decoder is based on U-Net. The implementation of conv block is shown in the following code;
\n\n (nn.Module):\n ():\n ().__init__()\n .conv1 = nn.Conv3d(in_channels, in_channels, kernel_size=, padding=, bias=, groups=in_channels)\n .norm1 = nn.GroupNorm(num_groups=, num_channels=in_channels)\n .act1 = nn.GELU()\n .conv2 = nn.Conv3d(in_channels, in_channels*, kernel_size=, bias=)\n .conv3 = nn.Conv3d(in_channels*, out_channels, kernel_size=, bias=)\n\n ():\n \n x = .conv1(x)\n x = .norm1(x)\n x = .conv2(x)\n x = .act1(x)\n x = .conv3(x)\n x\nAnd the implementation of the decoder is shown in the following code;
\n (nn.Module):\n ():\n ().__init__()\n .up3 = nn.Upsample(scale_factor=(,,), mode=, align_corners=)\n \n .dec3 = ConvBlock3D(in_channels=encoder_dims[] + encoder_dims[], out_channels=decoder_dims[])\n ...\n\n ():\n x, f0, f1, f2, f3 = features\n\n \n d3 = .up3(f3)\n d3 = torch.cat([d3, f2], dim=)\n d3 = .dec3(d3)\n ...\nSince the ground truth for segmentation is simply a sphere, I used a basic nn.Upsample for upsampling instead of nn.ConvTranspose3d to reduce parameters.
():\n lower, upper = np.percentile(x, (, ))\n x = np.clip(x, lower, upper)\n x = x - np.(x)\n x = x / np.(x)\n x\ninference code: https://www.kaggle.com/code/wadakoki/czii-9th-inference
\ntraining code: https://colab.research.google.com/drive/1OG5Rap9zUkKG6fnGZ8kP0JwwFZfSTgCc?usp=sharing
\nmodel checkpoint: https://www.kaggle.com/datasets/wadakoki/czii-9th-final-models/data
\ngenerated mask: https://www.kaggle.com/datasets/wadakoki/czii-mask-small
\ncode for mask generation: https://www.kaggle.com/code/wadakoki/czii-gen-small-masks
Requirements for training code on google colab
\nFirst of all, I would like to express my sincere gratitude to the competition host and the Kaggle staff for organizing such a fascinating competition. I thoroughly enjoyed this competition and learned a great deal in the process!
\nFurthermore, I'd like to thank @hengck23 and @davidlist . @hengck23 's discussion and notebook were the starting point for my solution, and Many of the discussions and comments from @davidlist were extremely helpful.
\nMy solution is simple and straightforward. I performed segmentation using a 3D ConvNeXt-like model and further employed an ensemble of as many models as possible. Subsequently, I computed the centroids of the particles using cc3d and curated the predicted centroids with DBSCAN.
\nI used ground truth masks with an adjusted radius for each particle. As a result, the public leaderboard score jumped 0.02~0.04. Concletely, I adjust mask size as shown in the below table;
\n| apo-ferritin(r=60) | \nbeta-galactosidase(r=90) | \nribosome(r=150) | \nthyroglobulin(r=130) | \nvirus-like-particle(r=135) | \n
|---|---|---|---|---|
| 60/2 | \n90/2 | \n150/3 | \n130/3 | \n135/3 | \n
Since the competition metric requires that the predicted centroid falls within a radius of r×0.5 from the ground truth centroid, I believe that multiplying the radius of each particle by a factor of 0.5 or less is reasonable.
\nIn this section, I explain my model architecture both encoder and decoder.
\nTo implement the encoder, I started with the ConvNeXt, since ConvNeXt is very powerful and fast. Then, I customized the model to suit the task that is to predict small particles. These changes from the original ConvNeXt are below;
\nThe flow of the decoder is based on U-Net. The implementation of conv block is shown in the following code;
\n\n (nn.Module):\n ():\n ().__init__()\n .conv1 = nn.Conv3d(in_channels, in_channels, kernel_size=, padding=, bias=, groups=in_channels)\n .norm1 = nn.GroupNorm(num_groups=, num_channels=in_channels)\n .act1 = nn.GELU()\n .conv2 = nn.Conv3d(in_channels, in_channels*, kernel_size=, bias=)\n .conv3 = nn.Conv3d(in_channels*, out_channels, kernel_size=, bias=)\n\n ():\n \n x = .conv1(x)\n x = .norm1(x)\n x = .conv2(x)\n x = .act1(x)\n x = .conv3(x)\n x\nAnd the implementation of the decoder is shown in the following code;
\n (nn.Module):\n ():\n ().__init__()\n .up3 = nn.Upsample(scale_factor=(,,), mode=, align_corners=)\n \n .dec3 = ConvBlock3D(in_channels=encoder_dims[] + encoder_dims[], out_channels=decoder_dims[])\n ...\n\n ():\n x, f0, f1, f2, f3 = features\n\n \n d3 = .up3(f3)\n d3 = torch.cat([d3, f2], dim=)\n d3 = .dec3(d3)\n ...\nSince the ground truth for segmentation is simply a sphere, I used a basic nn.Upsample for upsampling instead of nn.ConvTranspose3d to reduce parameters.
():\n lower, upper = np.percentile(x, (, ))\n x = np.clip(x, lower, upper)\n x = x - np.(x)\n x = x / np.(x)\n x\ninference code: https://www.kaggle.com/code/wadakoki/czii-9th-inference
\ntraining code: https://colab.research.google.com/drive/1OG5Rap9zUkKG6fnGZ8kP0JwwFZfSTgCc?usp=sharing
\nmodel checkpoint: https://www.kaggle.com/datasets/wadakoki/czii-9th-final-models/data
\ngenerated mask: https://www.kaggle.com/datasets/wadakoki/czii-mask-small
\ncode for mask generation: https://www.kaggle.com/code/wadakoki/czii-gen-small-masks
Requirements for training code on google colab
\n@wadakoki Thanks for the shout out! Congratulations!!!
", "rawMarkdown": "wadakoki Thanks for the shout out! Congratulations!!!", "votes": null }, { "id": "3116771", "postDate": "02/06/2025 09:42:39", "content": "Wow, a great implementation with such important details! Congratulations on your solo gold and prize!
", "rawMarkdown": "Wow, a great implementation with such important details! Congratulations on your solo gold and prize!", "votes": null }, { "id": "3118807", "postDate": "02/08/2025 14:07:31", "content": "[EDIT] I have added the training code, inference code, and the datasets used in this competition to the solution.
", "rawMarkdown": "[EDIT] I have added the training code, inference code, and the datasets used in this competition to the solution.", "votes": null } ], "comments": [ { "id": 3116652, "author_name": "davidlist", "author_url": "", "post_date": "02/06/2025 06:55:40", "content": "@wadakoki Thanks for the shout out! Congratulations!!!
", "votes": null, "replies": [] }, { "id": 3116771, "author_name": "snnclsr", "author_url": "", "post_date": "02/06/2025 09:42:39", "content": "Wow, a great implementation with such important details! Congratulations on your solo gold and prize!
", "votes": null, "replies": [] }, { "id": 3118807, "author_name": "wadakoki", "author_url": "", "post_date": "02/08/2025 14:07:31", "content": "[EDIT] I have added the training code, inference code, and the datasets used in this competition to the solution.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3116563": "First of all, I would like to express my sincere gratitude to the competition host and the Kaggle staff for organizing such a fascinating competition. I thoroughly enjoyed this competition and learned a great deal in the process!\n\nFurthermore, I'd like to thank @hengck23 and @davidlist . @hengck23 's discussion and notebook were the starting point for my solution, and Many of the discussions and comments from @davidlist were extremely helpful.\n\n## Summary\n\nMy solution is simple and straightforward. I performed segmentation using a 3D ConvNeXt-like model and further employed an ensemble of as many models as possible. Subsequently, I computed the centroids of the particles using cc3d and curated the predicted centroids with DBSCAN. \n\n## Segmentation Mask\n\nI used ground truth masks with an adjusted radius for each particle. As a result, the public leaderboard score jumped 0.02~0.04. Concletely, I adjust mask size as shown in the below table; \n\n| apo-ferritin(r=60) | beta-galactosidase(r=90) | ribosome(r=150) | thyroglobulin(r=130) | virus-like-particle(r=135) |\n| --- | --- | --- | --- | --- |\n| 60/2 | 90/2 | 150/3 | 130/3 | 135/3 |\n\nSince the competition metric requires that the predicted centroid falls within a radius of r×0.5 from the ground truth centroid, I believe that multiplying the radius of each particle by a factor of 0.5 or less is reasonable. \n\n## Model Architecture\n\nIn this section, I explain my model architecture both encoder and decoder. \n\n### Encoder\n\nTo implement the encoder, I started with the ConvNeXt, since ConvNeXt is very powerful and fast. Then, I customized the model to suit the task that is to predict small particles. These changes from the original ConvNeXt are below;\n\n- 2D to 3D for all convolutional layer\n- 4x4 stem -> 2x2 stem. Because the particles are very small, I believed that a large stem would adversely affect the model's predictions, especially apo-ferritin and beta-galactosidase. This modification slowed down the model's inference, but it improved its cross-validation performance. \n- 7x7 kernel size in conv block ->3x3. The reason why I modified like this is the same as the stem. This modification improved both inference speed and cv. \n- (3, 3, 9, 3) number of block -> (3, 3, 3, 3). this modification is to reduce a number of parameters and get more inference speed. cv did not decrease. \n\n### Decoder\n\nThe flow of the decoder is based on U-Net. The implementation of conv block is shown in the following code; \n\n```python\n# conv block for decoder\nclass ConvBlock3D(nn.Module):\n def __init__(self, in_channels, out_channels):\n super().__init__()\n self.conv1 = nn.Conv3d(in_channels, in_channels, kernel_size=3, padding=1, bias=False, groups=in_channels)\n self.norm1 = nn.GroupNorm(num_groups=1, num_channels=in_channels)\n self.act1 = nn.GELU()\n self.conv2 = nn.Conv3d(in_channels, in_channels*3, kernel_size=1, bias=False)\n self.conv3 = nn.Conv3d(in_channels*3, out_channels, kernel_size=1, bias=False)\n\n def forward(self, x):\n # x: (B, C, D, H, W)\n x = self.conv1(x)\n x = self.norm1(x)\n x = self.conv2(x)\n x = self.act1(x)\n x = self.conv3(x)\n return x\n```\nAnd the implementation of the decoder is shown in the following code; \n\n```python\nclass Decoder3D(nn.Module):\n def __init__(self,\n encoder_dims=[64, 128, 256, 512],\n decoder_dims=[32, 64, 128, 256],\n out_channels=5):\n super().__init__()\n self.up3 = nn.Upsample(scale_factor=(2,2,2), mode='trilinear', align_corners=True)\n #self.up3 = nn.ConvTranspose3d(encoder_dims[3], encoder_dims[3], kernel_size=2, stride=2)\n self.dec3 = ConvBlock3D(in_channels=encoder_dims[3] + encoder_dims[2], out_channels=decoder_dims[2])\n ...\n\n def forward(self, features):\n x, f0, f1, f2, f3 = features\n\n # --- 1) f3 -> f2 ---\n d3 = self.up3(f3)\n d3 = torch.cat([d3, f2], dim=1)\n d3 = self.dec3(d3)\n ...\n```\n\nSince the ground truth for segmentation is simply a sphere, I used a basic `nn.Upsample` for upsampling instead of `nn.ConvTranspose3d` to reduce parameters. \n\n## Other Details\n\n- loss dunction is bce\n- My inference code is based on @hengck 's [this notebook](https://www.kaggle.com/code/hengck23/1-hr-fast-2d-3d-unet-resnet34d-scanner-tta) and the idea of DBScan came from @linheshen 's [notebook](https://www.kaggle.com/code/hengck23/1-hr-fast-2d-3d-unet-resnet34d-scanner-tta). \n- The window size for inference is (32, 320, 320) (overwrap of z axis is 6). The window size for training is (32, 256, 256). \n- TTAs are rot90, 180 and 270. \n- preprocessing is only normalization\n```python\n def normalize_numpy(self, x):\n lower, upper = np.percentile(x, (1, 99))\n x = np.clip(x, lower, upper)\n x = x - np.min(x)\n x = x / np.max(x)\n return x\n```\n- Augmentations for training are rot90, 180, 270, and xyz flip. \n\n## What did not work\n\n- pretraining with synthetic data provided by host\n- more and more smaller mask(e.g. factor=0.25 for large particles and factor=0.4 for small particles)\n- flip tta instead of rot90\n\n## Code & Model Checkpoint\ninference code: https://www.kaggle.com/code/wadakoki/czii-9th-inference\ntraining code: https://colab.research.google.com/drive/1OG5Rap9zUkKG6fnGZ8kP0JwwFZfSTgCc?usp=sharing\nmodel checkpoint: https://www.kaggle.com/datasets/wadakoki/czii-9th-final-models/data\ngenerated mask: https://www.kaggle.com/datasets/wadakoki/czii-mask-small\ncode for mask generation: https://www.kaggle.com/code/wadakoki/czii-gen-small-masks\n\n**Requirements for training code on google colab**\n- GPU: L4\n- RAM: High Memory", "3116652": "wadakoki Thanks for the shout out! Congratulations!!!", "3116771": "Wow, a great implementation with such important details! Congratulations on your solo gold and prize!", "3118807": "[EDIT] I have added the training code, inference code, and the datasets used in this competition to the solution." }, "source": "meta" }