{ "id": 586057, "title": "2nd Place Solution – Global AI Hackathon 2025 by Elucidata", "url": "/competitions/el-hackathon-2025/discussion/586057", "author_name": "", "post_date": "2025-06-24T19:34:12.680169700Z", "votes": 3, "comment_count": 8, "views": 0, "content": "

Hi everyone!
\nI’m excited to share my 2nd place winning solution for the Global AI Hackathon 2025 by Elucidata.

\n

Highlights of my approach:

\n

Stain normalization using Vahadane’s method with staintools

\n

Feature extraction with ConvNeXt-Small

\n

Spatial fusion with coordinate encodings

\n

Custom MSE + differentiable Spearman loss

\n

TTA + Gaussian smoothing in postprocessing****

\n

Code & Details:
\nYou can find the full pipeline, including preprocessing, model, and inference, in my GitHub repo:
\nhttps://github.com/Masry5/Global-AI-Hackathon-25-by-Elucidata

\n

Let me know your thoughts or if you have questions.

", "messages": [ { "id": "3231709", "postDate": "06/24/2025 19:34:12", "content": "

Hi everyone!
\nI’m excited to share my 2nd place winning solution for the Global AI Hackathon 2025 by Elucidata.

\n

Highlights of my approach:

\n

Stain normalization using Vahadane’s method with staintools

\n

Feature extraction with ConvNeXt-Small

\n

Spatial fusion with coordinate encodings

\n

Custom MSE + differentiable Spearman loss

\n

TTA + Gaussian smoothing in postprocessing****

\n

Code & Details:
\nYou can find the full pipeline, including preprocessing, model, and inference, in my GitHub repo:
\nhttps://github.com/Masry5/Global-AI-Hackathon-25-by-Elucidata

\n

Let me know your thoughts or if you have questions.

", "rawMarkdown": "**Hi everyone!**\nI’m excited to share my 2nd place winning solution for the Global AI Hackathon 2025 by Elucidata.\n\n \n**Highlights of my approach:**\n\nStain normalization using Vahadane’s method with staintools\n\nFeature extraction with ConvNeXt-Small\n\nSpatial fusion with coordinate encodings\n\nCustom MSE + differentiable Spearman loss\n\nTTA + Gaussian smoothing in postprocessing****\n\n\n****Code & Details:****\nYou can find the full pipeline, including preprocessing, model, and inference, in my GitHub repo:\nhttps://github.com/Masry5/Global-AI-Hackathon-25-by-Elucidata\n\n\nLet me know your thoughts or if you have questions.", "votes": null }, { "id": "3237611", "postDate": "07/01/2025 07:33:35", "content": "

Using the same image but different regions could be a great strategy. Thanks for your sharing.

", "rawMarkdown": "Using the same image but different regions could be a great strategy. Thanks for your sharing.", "votes": null }, { "id": "3237621", "postDate": "07/01/2025 07:41:14", "content": "

Also, It's super surprised that your train vs validation plot is almost reflect the model apply in the new image, which means this method can solve the domain-shift problem. It's a really interested observation. good job.

", "rawMarkdown": "Also, It's super surprised that your train vs validation plot is almost reflect the model apply in the new image, which means this method can solve the domain-shift problem. It's a really interested observation. good job.", "votes": null }, { "id": "3239651", "postDate": "07/03/2025 02:59:50", "content": "

Just for my curiosity. Would I know your highest private score?

", "rawMarkdown": "Just for my curiosity. Would I know your highest private score?", "votes": null }, { "id": "3241942", "postDate": "07/05/2025 11:26:36", "content": "

0.725

", "rawMarkdown": "0.725", "votes": null }, { "id": "3242153", "postDate": "07/05/2025 16:27:02", "content": "

Thanks for answering! I'm also wondering did you try to remove the Spatial fusion with coordinate encodings? In my case, the spatial fusion doesn't work really well.

", "rawMarkdown": "Thanks for answering! I'm also wondering did you try to remove the Spatial fusion with coordinate encodings? In my case, the spatial fusion doesn't work really well.", "votes": null }, { "id": "3243038", "postDate": "07/06/2025 17:45:43", "content": "

Yes, I did try versions without the spatial fusion , but I actually got better results when I included it.
\nIn my case, the spatial fusion seemed to help the model capture spatial context better,

", "rawMarkdown": "Yes, I did try versions without the spatial fusion , but I actually got better results when I included it.\nIn my case, the spatial fusion seemed to help the model capture spatial context better,", "votes": null }, { "id": "3244460", "postDate": "07/08/2025 07:07:00", "content": "

that's an interesting finding. Thanks! I guess is becasue you didn't preprocess the expression and in this case the location encoding become meaningful.

", "rawMarkdown": "that's an interesting finding. Thanks! I guess is becasue you didn't preprocess the expression and in this case the location encoding become meaningful.", "votes": null }, { "id": "3246670", "postDate": "07/11/2025 10:19:01", "content": "

Well done! Your solution suggests that ConvNeXt’s transformer-inspired design, when paired with spatial fusion was key to modeling the rich spatial structure for this histology data. Thanks for sharing.

", "rawMarkdown": "Well done! Your solution suggests that ConvNeXt’s transformer-inspired design, when paired with spatial fusion was key to modeling the rich spatial structure for this histology data. Thanks for sharing.", "votes": null } ], "comments": [ { "id": 3237611, "author_name": "dingyangwang", "author_url": "", "post_date": "07/01/2025 07:33:35", "content": "

Using the same image but different regions could be a great strategy. Thanks for your sharing.

", "votes": null, "replies": [ { "id": 3237621, "author_name": "dingyangwang", "author_url": "", "post_date": "07/01/2025 07:41:14", "content": "

Also, It's super surprised that your train vs validation plot is almost reflect the model apply in the new image, which means this method can solve the domain-shift problem. It's a really interested observation. good job.

", "votes": null, "replies": [] } ] }, { "id": 3239651, "author_name": "dingyangwang", "author_url": "", "post_date": "07/03/2025 02:59:50", "content": "

Just for my curiosity. Would I know your highest private score?

", "votes": null, "replies": [ { "id": 3241942, "author_name": "masry1", "author_url": "", "post_date": "07/05/2025 11:26:36", "content": "

0.725

", "votes": null, "replies": [ { "id": 3242153, "author_name": "dingyangwang", "author_url": "", "post_date": "07/05/2025 16:27:02", "content": "

Thanks for answering! I'm also wondering did you try to remove the Spatial fusion with coordinate encodings? In my case, the spatial fusion doesn't work really well.

", "votes": null, "replies": [ { "id": 3243038, "author_name": "masry1", "author_url": "", "post_date": "07/06/2025 17:45:43", "content": "

Yes, I did try versions without the spatial fusion , but I actually got better results when I included it.
\nIn my case, the spatial fusion seemed to help the model capture spatial context better,

", "votes": null, "replies": [ { "id": 3244460, "author_name": "dingyangwang", "author_url": "", "post_date": "07/08/2025 07:07:00", "content": "

that's an interesting finding. Thanks! I guess is becasue you didn't preprocess the expression and in this case the location encoding become meaningful.

", "votes": null, "replies": [] } ] } ] } ] } ] }, { "id": 3246670, "author_name": "deepaksaldanha", "author_url": "", "post_date": "07/11/2025 10:19:01", "content": "

Well done! Your solution suggests that ConvNeXt’s transformer-inspired design, when paired with spatial fusion was key to modeling the rich spatial structure for this histology data. Thanks for sharing.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3231709": "**Hi everyone!**\nI’m excited to share my 2nd place winning solution for the Global AI Hackathon 2025 by Elucidata.\n\n \n**Highlights of my approach:**\n\nStain normalization using Vahadane’s method with staintools\n\nFeature extraction with ConvNeXt-Small\n\nSpatial fusion with coordinate encodings\n\nCustom MSE + differentiable Spearman loss\n\nTTA + Gaussian smoothing in postprocessing****\n\n\n****Code & Details:****\nYou can find the full pipeline, including preprocessing, model, and inference, in my GitHub repo:\nhttps://github.com/Masry5/Global-AI-Hackathon-25-by-Elucidata\n\n\nLet me know your thoughts or if you have questions.", "3237611": "Using the same image but different regions could be a great strategy. Thanks for your sharing.", "3237621": "Also, It's super surprised that your train vs validation plot is almost reflect the model apply in the new image, which means this method can solve the domain-shift problem. It's a really interested observation. good job.", "3239651": "Just for my curiosity. Would I know your highest private score?", "3241942": "0.725", "3242153": "Thanks for answering! I'm also wondering did you try to remove the Spatial fusion with coordinate encodings? In my case, the spatial fusion doesn't work really well.", "3243038": "Yes, I did try versions without the spatial fusion , but I actually got better results when I included it.\nIn my case, the spatial fusion seemed to help the model capture spatial context better,", "3244460": "that's an interesting finding. Thanks! I guess is becasue you didn't preprocess the expression and in this case the location encoding become meaningful.", "3246670": "Well done! Your solution suggests that ConvNeXt’s transformer-inspired design, when paired with spatial fusion was key to modeling the rich spatial structure for this histology data. Thanks for sharing." }, "source": "meta" }