{
  "id": 280121,
  "title": "2d vs 3d cnn results",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/280121",
  "author_name": "imori",
  "post_date": "2021-10-20T15:10:23.097000",
  "votes": 7,
  "comment_count": 0,
  "views": 0,
  "content": "<p>This is interesting competition!</p>\n<p>I tried some approached for 3d.</p>\n<h1>Summary</h1>\n<table>\n<thead>\n<tr>\n<th>No</th>\n<th>method</th>\n<th>LB</th>\n<th>PB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>1</td>\n<td>2d vit + slice prediction + max probability</td>\n<td>0.52484</td>\n<td>0.55983</td>\n</tr>\n<tr>\n<td>2</td>\n<td>2d vit + slice prediction + mean probability</td>\n<td>0.48837</td>\n<td>0.53794</td>\n</tr>\n<tr>\n<td>3</td>\n<td>3d resnet50 + 4 type image concat</td>\n<td>0.65221</td>\n<td>0.50628</td>\n</tr>\n<tr>\n<td>4</td>\n<td>3d resnet50 x 4 type image modeling</td>\n<td>0.64587</td>\n<td>0.50463</td>\n</tr>\n</tbody>\n</table>\n<p>LB vs PB is very mysterious.</p>\n<p>I tried image embedding using metric learning to visualize similarity, but there are similar embeddings between train and public data. However, PB looks different from LB. Probably, my embedding approach is not good, but I think there are some domain shift.</p>\n<h2>2d approach</h2>\n<p>I trained ViT using slices with label in train_labels.csv. When inference, I predicted each image using 5 fold ViT and get mean score of each image. Finally, I got max(No.1) or mean(No.2) prediction value using all slices for one BraTS21ID.</p>\n<h2>3d approach</h2>\n<p>I trained 3d-resnet50 using sequence image datas with 128 samples.<br>\nInput is 4 dim (FLAIR, T1w, T1wCE, T2w) or 1dim x 4 models. (No.3, 4).<br>\nWhen  inference, I predicted using 3d model with same input image.</p>\n<h2>Pre-processing</h2>\n<p>Original image has large 0 zone, so I cropped brain position and ignored some images without brain. However, this didn't help cv score.</p>\n<h2>Thoughts</h2>\n<p>I could not find key tricks, but tried some 3d approaches for 2 weeks. I had shake down, but this was interesting competition!</p>",
  "messages": [
    {
      "id": 1551386,
      "postDate": "2021-10-20T15:10:23.097Z",
      "content": "<p>This is interesting competition!</p>\n<p>I tried some approached for 3d.</p>\n<h1>Summary</h1>\n<table>\n<thead>\n<tr>\n<th>No</th>\n<th>method</th>\n<th>LB</th>\n<th>PB</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>1</td>\n<td>2d vit + slice prediction + max probability</td>\n<td>0.52484</td>\n<td>0.55983</td>\n</tr>\n<tr>\n<td>2</td>\n<td>2d vit + slice prediction + mean probability</td>\n<td>0.48837</td>\n<td>0.53794</td>\n</tr>\n<tr>\n<td>3</td>\n<td>3d resnet50 + 4 type image concat</td>\n<td>0.65221</td>\n<td>0.50628</td>\n</tr>\n<tr>\n<td>4</td>\n<td>3d resnet50 x 4 type image modeling</td>\n<td>0.64587</td>\n<td>0.50463</td>\n</tr>\n</tbody>\n</table>\n<p>LB vs PB is very mysterious.</p>\n<p>I tried image embedding using metric learning to visualize similarity, but there are similar embeddings between train and public data. However, PB looks different from LB. Probably, my embedding approach is not good, but I think there are some domain shift.</p>\n<h2>2d approach</h2>\n<p>I trained ViT using slices with label in train_labels.csv. When inference, I predicted each image using 5 fold ViT and get mean score of each image. Finally, I got max(No.1) or mean(No.2) prediction value using all slices for one BraTS21ID.</p>\n<h2>3d approach</h2>\n<p>I trained 3d-resnet50 using sequence image datas with 128 samples.<br>\nInput is 4 dim (FLAIR, T1w, T1wCE, T2w) or 1dim x 4 models. (No.3, 4).<br>\nWhen  inference, I predicted using 3d model with same input image.</p>\n<h2>Pre-processing</h2>\n<p>Original image has large 0 zone, so I cropped brain position and ignored some images without brain. However, this didn't help cv score.</p>\n<h2>Thoughts</h2>\n<p>I could not find key tricks, but tried some 3d approaches for 2 weeks. I had shake down, but this was interesting competition!</p>",
      "rawMarkdown": "This is interesting competition!\n\nI tried some approached for 3d.\n\n# Summary\n\n|No|method|LB|PB|\n|:---:|:---:|:---:|:---:|\n|1| 2d vit + slice prediction + max probability | 0.52484 | 0.55983 |\n|2| 2d vit + slice prediction + mean probability | 0.48837 | 0.53794 |\n|3| 3d resnet50 + 4 type image concat | 0.65221 | 0.50628 |\n|4| 3d resnet50 x 4 type image modeling | 0.64587 | 0.50463 |\n\nLB vs PB is very mysterious.\n\nI tried image embedding using metric learning to visualize similarity, but there are similar embeddings between train and public data. However, PB looks different from LB. Probably, my embedding approach is not good, but I think there are some domain shift.\n\n\n## 2d approach\n\nI trained ViT using slices with label in train_labels.csv. When inference, I predicted each image using 5 fold ViT and get mean score of each image. Finally, I got max(No.1) or mean(No.2) prediction value using all slices for one BraTS21ID.\n\n## 3d approach\n\nI trained 3d-resnet50 using sequence image datas with 128 samples.\nInput is 4 dim (FLAIR, T1w, T1wCE, T2w) or 1dim x 4 models. (No.3, 4).\nWhen  inference, I predicted using 3d model with same input image.\n\n## Pre-processing\n\nOriginal image has large 0 zone, so I cropped brain position and ignored some images without brain. However, this didn't help cv score.\n\n## Thoughts\n\nI could not find key tricks, but tried some 3d approaches for 2 weeks. I had shake down, but this was interesting competition!\n\n",
      "votes": 7
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1551386": "This is interesting competition!\n\nI tried some approached for 3d.\n\n# Summary\n\n|No|method|LB|PB|\n|:---:|:---:|:---:|:---:|\n|1| 2d vit + slice prediction + max probability | 0.52484 | 0.55983 |\n|2| 2d vit + slice prediction + mean probability | 0.48837 | 0.53794 |\n|3| 3d resnet50 + 4 type image concat | 0.65221 | 0.50628 |\n|4| 3d resnet50 x 4 type image modeling | 0.64587 | 0.50463 |\n\nLB vs PB is very mysterious.\n\nI tried image embedding using metric learning to visualize similarity, but there are similar embeddings between train and public data. However, PB looks different from LB. Probably, my embedding approach is not good, but I think there are some domain shift.\n\n\n## 2d approach\n\nI trained ViT using slices with label in train_labels.csv. When inference, I predicted each image using 5 fold ViT and get mean score of each image. Finally, I got max(No.1) or mean(No.2) prediction value using all slices for one BraTS21ID.\n\n## 3d approach\n\nI trained 3d-resnet50 using sequence image datas with 128 samples.\nInput is 4 dim (FLAIR, T1w, T1wCE, T2w) or 1dim x 4 models. (No.3, 4).\nWhen  inference, I predicted using 3d model with same input image.\n\n## Pre-processing\n\nOriginal image has large 0 zone, so I cropped brain position and ignored some images without brain. However, this didn't help cv score.\n\n## Thoughts\n\nI could not find key tricks, but tried some 3d approaches for 2 weeks. I had shake down, but this was interesting competition!\n\n"
  }
}