The size of the image data gives rise to a compensation strategy to prevent running out of RAM. I saw some approaches that create datasets with image slices and I wonder if we then can even say whether one particular slice contains the information we need for the classification. In my understanding (e.g. this paper), there might be areas that are more crucial for the inference than others. How would you label such a dataset? Did I miss an important workaround that everyone uses?
\nAny thoughts or ideas are appreciated.
The size of the image data gives rise to a compensation strategy to prevent running out of RAM. I saw some approaches that create datasets with image slices and I wonder if we then can even say whether one particular slice contains the information we need for the classification. In my understanding (e.g. this paper), there might be areas that are more crucial for the inference than others. How would you label such a dataset? Did I miss an important workaround that everyone uses?
\nAny thoughts or ideas are appreciated.
The concept of Multiple Instance Learning seems to be applicable by grouping slices into bags and assigning the label to the bag instead of each individual slice. Here is an interesting Keras example. Not sure if it works for this case but worth experimenting in my opinion. Hope it helps!
", "rawMarkdown": "The concept of [Multiple Instance Learning](https://en.wikipedia.org/wiki/Multiple_instance_learning) seems to be applicable by grouping slices into bags and assigning the label to the bag instead of each individual slice. Here is an interesting Keras [example](https://keras.io/examples/vision/attention_mil_classification/). Not sure if it works for this case but worth experimenting in my opinion. Hope it helps!", "votes": 2 }, { "id": 1857935, "postDate": "2022-07-16T14:18:20.717Z", "content": "A very good question, first of all, even if we tile a slide, it is true that not all the tiles are going to have a signal which tells us which class it belongs to, there is no workaround that is public on kaggle to my knowledge, but that does not mean that it can not be done…
", "rawMarkdown": "A very good question, first of all, even if we tile a slide, it is true that not all the tiles are going to have a signal which tells us which class it belongs to, there is no workaround that is public on kaggle to my knowledge, but that does not mean that it can not be done..." } ], "comments": [ { "id": 1858169, "author_name": "twoisprime", "author_url": "", "post_date": "2022-07-16T18:03:21.537000", "content": "The concept of Multiple Instance Learning seems to be applicable by grouping slices into bags and assigning the label to the bag instead of each individual slice. Here is an interesting Keras example. Not sure if it works for this case but worth experimenting in my opinion. Hope it helps!
", "votes": 2, "replies": [] }, { "id": 1857935, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2022-07-16T14:18:20.717000", "content": "A very good question, first of all, even if we tile a slide, it is true that not all the tiles are going to have a signal which tells us which class it belongs to, there is no workaround that is public on kaggle to my knowledge, but that does not mean that it can not be done…
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1857811": "The size of the image data gives rise to a compensation strategy to prevent running out of RAM. I saw some approaches that create datasets with image slices and I wonder if we then can even say whether one particular slice contains the information we need for the classification. In my understanding (e.g. [this paper](https://doi.org/10.3324/haematol.2019.219881)), there might be areas that are more crucial for the inference than others. How would you label such a dataset? Did I miss an important workaround that everyone uses?\nAny thoughts or ideas are appreciated.", "1858169": "The concept of [Multiple Instance Learning](https://en.wikipedia.org/wiki/Multiple_instance_learning) seems to be applicable by grouping slices into bags and assigning the label to the bag instead of each individual slice. Here is an interesting Keras [example](https://keras.io/examples/vision/attention_mil_classification/). Not sure if it works for this case but worth experimenting in my opinion. Hope it helps!", "1857935": "A very good question, first of all, even if we tile a slide, it is true that not all the tiles are going to have a signal which tells us which class it belongs to, there is no workaround that is public on kaggle to my knowledge, but that does not mean that it can not be done..." } }