{ "id": 599564, "title": "Are we allowed to separate modality in test (loop all test cases and separate them first) and apply modality-specific models ?", "url": "/competitions/rsna-intracranial-aneurysm-detection/discussion/599564", "author_name": "Leo Yang", "post_date": "2025-08-17T11:30:23.965000", "votes": 1, "comment_count": 6, "views": 0, "content": "

Hi everyone,

\n

I noticed that the test metadata includes modality tags 'CT' and 'MR'. Given this, are we allowed to separate the test serise by modality and apply different models for CT And MR respectively ?

\n

Best
\nLeo

", "messages": [ { "id": 3270815, "postDate": "2025-08-17T12:00:49.267Z", "content": "

Of course not. That's cheating! Just kidding, do whathever you think will help and run in less than 12 hours. But, you don't need to double loop, you can load both models and check each sample modality. Load volumes is time consuming. And consider you won't have any explicit information about T1/T2 subsets in MRI… Personally for what I've seen till now I think is not hard for models generalize all those type of data. So Id' say use a single model.

", "rawMarkdown": "Of course not. That's cheating! Just kidding, do whathever you think will help and run in less than 12 hours. But, you don't need to double loop, you can load both models and check each sample modality. Load volumes is time consuming. And consider you won't have any explicit information about T1/T2 subsets in MRI... Personally for what I've seen till now I think is not hard for models generalize all those type of data. So Id' say use a single model.", "votes": 1, "replies": [ { "id": 3271062, "postDate": "2025-08-18T00:35:44.087Z", "content": "

But we need to a make a batch (I assume batch size should be large enough to speed up the test) before feeding them into the model. How to ensure each batch only contains one modality without a double loop ?

", "rawMarkdown": "But we need to a make a batch (I assume batch size should be large enough to speed up the test) before feeding them into the model. How to ensure each batch only contains one modality without a double loop ?", "replies": [ { "id": 3271068, "postDate": "2025-08-18T01:20:27.270Z", "content": "

Then yes, you should double loop. Or train a single general model. I must confess I haven't reached that point yet. But at least for segmentation, no specific models are needed. And since diagnostic is based on brain arteries shape, should also be possible a single general model for that.

", "rawMarkdown": "Then yes, you should double loop. Or train a single general model. I must confess I haven't reached that point yet. But at least for segmentation, no specific models are needed. And since diagnostic is based on brain arteries shape, should also be possible a single general model for that." } ] } ] }, { "id": 3270809, "postDate": "2025-08-17T11:30:23.967Z", "content": "

Hi everyone,

\n

I noticed that the test metadata includes modality tags 'CT' and 'MR'. Given this, are we allowed to separate the test serise by modality and apply different models for CT And MR respectively ?

\n

Best
\nLeo

", "rawMarkdown": "Hi everyone,\n\nI noticed that the test metadata includes modality tags 'CT' and 'MR'. Given this, are we allowed to separate the test serise by modality and apply different models for CT And MR respectively ?\n\nBest \nLeo", "votes": 1 }, { "id": 3270875, "postDate": "2025-08-17T14:28:27.697Z", "content": "

u can load them separately by ds.modality， and if modality==ct do ct_model(ct) if modality==mr do mr_model(mr)

", "rawMarkdown": "u can load them separately by ds.modality， and if modality==ct do ct_model(ct) if modality==mr do mr_model(mr)", "votes": -1 }, { "id": 3271406, "postDate": "2025-08-18T19:12:03.810Z", "content": "

My understanding is that your model will only be allowed to see each test set case once during model testing. This is essential for preventing unwanted probing of the test set.

", "rawMarkdown": "My understanding is that your model will only be allowed to see each test set case once during model testing. This is essential for preventing unwanted probing of the test set.", "replies": [ { "id": 3271495, "postDate": "2025-08-19T02:48:45.043Z", "content": "

Oh, thanks for clarification!

", "rawMarkdown": "Oh, thanks for clarification!" } ] } ], "comments": [ { "id": 3270815, "author_name": "Ángel Jacinto Sánchez Ruiz", "author_url": "", "post_date": "2025-08-17T12:00:49.267000", "content": "

Of course not. That's cheating! Just kidding, do whathever you think will help and run in less than 12 hours. But, you don't need to double loop, you can load both models and check each sample modality. Load volumes is time consuming. And consider you won't have any explicit information about T1/T2 subsets in MRI… Personally for what I've seen till now I think is not hard for models generalize all those type of data. So Id' say use a single model.

", "votes": 1, "replies": [ { "id": 3271062, "author_name": "Leo Yang", "author_url": "", "post_date": "2025-08-18T00:35:44.087000", "content": "

But we need to a make a batch (I assume batch size should be large enough to speed up the test) before feeding them into the model. How to ensure each batch only contains one modality without a double loop ?

", "votes": 0, "replies": [ { "id": 3271068, "author_name": "Ángel Jacinto Sánchez Ruiz", "author_url": "", "post_date": "2025-08-18T01:20:27.270000", "content": "

Then yes, you should double loop. Or train a single general model. I must confess I haven't reached that point yet. But at least for segmentation, no specific models are needed. And since diagnostic is based on brain arteries shape, should also be possible a single general model for that.

", "votes": 0, "replies": [] } ] } ] }, { "id": 3270875, "author_name": "Seeing Times", "author_url": "", "post_date": "2025-08-17T14:28:27.697000", "content": "

u can load them separately by ds.modality， and if modality==ct do ct_model(ct) if modality==mr do mr_model(mr)

", "votes": -1, "replies": [] }, { "id": 3271406, "author_name": "Evan Calabrese", "author_url": "", "post_date": "2025-08-18T19:12:03.810000", "content": "

My understanding is that your model will only be allowed to see each test set case once during model testing. This is essential for preventing unwanted probing of the test set.

", "votes": 0, "replies": [ { "id": 3271495, "author_name": "Leo Yang", "author_url": "", "post_date": "2025-08-19T02:48:45.043000", "content": "

Oh, thanks for clarification!

", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "3270815": "Of course not. That's cheating! Just kidding, do whathever you think will help and run in less than 12 hours. But, you don't need to double loop, you can load both models and check each sample modality. Load volumes is time consuming. And consider you won't have any explicit information about T1/T2 subsets in MRI... Personally for what I've seen till now I think is not hard for models generalize all those type of data. So Id' say use a single model.", "3270809": "Hi everyone,\n\nI noticed that the test metadata includes modality tags 'CT' and 'MR'. Given this, are we allowed to separate the test serise by modality and apply different models for CT And MR respectively ?\n\nBest \nLeo", "3270875": "u can load them separately by ds.modality， and if modality==ct do ct_model(ct) if modality==mr do mr_model(mr)", "3271406": "My understanding is that your model will only be allowed to see each test set case once during model testing. This is essential for preventing unwanted probing of the test set." } }