{ "id": 448414, "title": "Overview of Published Research in Ovarian Cancer Subtyping AI ", "url": "/competitions/UBC-OCEAN/discussion/448414", "author_name": "JackBreen", "post_date": "2023-10-19T15:26:15.061000", "votes": 31, "comment_count": 2, "views": 0, "content": "

Hello all,
\nWe recently published a systematic review of AI for ovarian cancer diagnosis, so here's some observations about ovarian cancer subtyping:

\n\n

Outside of the review there have been some other potentially relevant papers:

\n\n

Please do comment to add any other relevant research, and check out the previous discussion posts providing resources on the biology of ovarian cancer subtypes, previous relevant challenges, and other relevant outlier detection work

", "messages": [ { "id": 2488914, "postDate": "2023-10-19T15:26:15.060Z", "content": "

Hello all,
\nWe recently published a systematic review of AI for ovarian cancer diagnosis, so here's some observations about ovarian cancer subtyping:

\n\n

Outside of the review there have been some other potentially relevant papers:

\n\n

Please do comment to add any other relevant research, and check out the previous discussion posts providing resources on the biology of ovarian cancer subtypes, previous relevant challenges, and other relevant outlier detection work

", "rawMarkdown": "Hello all,\nWe recently published a [systematic review of AI for ovarian cancer diagnosis](https://www.nature.com/articles/s41698-023-00432-6), so here's some observations about ovarian cancer subtyping:\n- There were only seven published studies on ovarian cancer subtyping, of which, three were older SVM-based approaches ([BenTaieb2015](https://link.springer.com/chapter/10.1007/978-3-319-24553-9_77), [BenTaieb2016](https://doi.org/10.4103/2153-3539.186899), [BenTaieb2017](https://doi.org/10.1016/j.media.2017.04.008)), and four were newer deep CNN approaches ([Levine2020](https://doi.org/10.1002/path.5509), [Boschman2022](https://doi.org/10.1002/path.5797), [Farahani2022](https://doi.org/10.1038/s41379-022-01146-z), [Kasture2022](https://doi.org/10.14445/22315381/IJETT-V70I3P235)).\n- All relevant studies relied on patch-level information (typically 256x256 to 512x512 pixel patches) due to the overwhelming size of whole slides. As such, multiple instance learning (MIL) methods were common.\n- The most directly relevant studies were both from UBC ([Boschman2022](https://doi.org/10.1002/path.5797), [Farahani2022](https://doi.org/10.1038/s41379-022-01146-z)), where this challenge was organised. These were the only studies in which hundreds of WSIs were classified at slide-level (not at patch-level). These studies reported AUCs well over 0.9. One of these studies has a [published code repository](https://github.com/AIMLab-UBC/ModernPath2022).\n- None of the previous studies used TMAs or had a class beside the five most common subtypes (except for non-cancer in [Kasture2022](https://doi.org/10.14445/22315381/IJETT-V70I3P235)), though there is a study on [high-grade serous vs serous borderline](https://doi.org/10.4103/jpi.jpi_76_20), and a number of studies on distinguishing tumour from benign tissue. None of the previous studies included outlier detection or uncertainty quantification. As such, this challenge is exploring uncharted territory.\n\nOutside of the review there have been some other potentially relevant papers:\n- Our paper where we found that in using whole slides for subtyping, it is typically sufficient to use only a fraction of the available information to reduce computational requirements (e.g. 10%) ([Breen2023](http://dx.doi.org/10.1117/12.2653869), [code](https://github.com/scjjb/DRAS-MIL)). \n- [Classification of Epithelial Ovarian Carcinoma Whole-Slide Pathology Images Using Deep Transfer Learning](https://openreview.net/forum?id=VXdQD8B307) An earlier short paper from UBC.\n- [Heram : Multi-Magnification Graph-Structured Whole Slide Image Representation](https://lrjconan.github.io/UBC-EECE571F-DL-Structures/assets/sample_reports_2022_W1/report_03.pdf), a multi-scale graph-based method from UBC which reports greater performance than their previous approaches, and which is currently under review.\n\nPlease do comment to add any other relevant research, and check out the previous [discussion](https://www.kaggle.com/competitions/UBC-OCEAN/discussion/445470) [posts](https://www.kaggle.com/competitions/UBC-OCEAN/discussion/445472) providing resources on the biology of ovarian cancer subtypes, previous relevant challenges, and other relevant outlier detection work", "votes": 31 }, { "id": 2544010, "postDate": "2023-11-30T14:27:10.227Z", "content": "

Adding our newest research, where we found that lower magnifications actually gave better performance than higher magnifications, with performance best around 5x (where the data in this challenge starts at 20x for WSIs and 40x for TMAs). These lower magnificiations also significantly reduced the computational requirements:
\nhttps://arxiv.org/abs/2311.13956

", "rawMarkdown": "Adding [our newest research](https://arxiv.org/abs/2311.13956), where we found that lower magnifications actually gave better performance than higher magnifications, with performance best around 5x (where the data in this challenge starts at 20x for WSIs and 40x for TMAs). These lower magnificiations also significantly reduced the computational requirements: \nhttps://arxiv.org/abs/2311.13956 ", "votes": 6 }, { "id": 2489016, "postDate": "2023-10-19T16:26:32.460Z", "content": "

Hi Jack Breen,
\nAmazing article published on Nature: \"Artificial intelligence in ovarian cancer histopathology: a systematic review\"
\nhttps://www.nature.com/articles/s41698-023-00432-6

", "rawMarkdown": "Hi Jack Breen,\nAmazing article published on Nature: \"Artificial intelligence in ovarian cancer histopathology: a systematic review\"\nhttps://www.nature.com/articles/s41698-023-00432-6", "votes": 6 } ], "comments": [ { "id": 2544010, "author_name": "JackBreen", "author_url": "", "post_date": "2023-11-30T14:27:10.227000", "content": "

Adding our newest research, where we found that lower magnifications actually gave better performance than higher magnifications, with performance best around 5x (where the data in this challenge starts at 20x for WSIs and 40x for TMAs). These lower magnificiations also significantly reduced the computational requirements:
\nhttps://arxiv.org/abs/2311.13956

", "votes": 6, "replies": [] }, { "id": 2489016, "author_name": "Marília Prata", "author_url": "", "post_date": "2023-10-19T16:26:32.460000", "content": "

Hi Jack Breen,
\nAmazing article published on Nature: \"Artificial intelligence in ovarian cancer histopathology: a systematic review\"
\nhttps://www.nature.com/articles/s41698-023-00432-6

", "votes": 6, "replies": [] } ], "raw_markdown_by_id": { "2488914": "Hello all,\nWe recently published a [systematic review of AI for ovarian cancer diagnosis](https://www.nature.com/articles/s41698-023-00432-6), so here's some observations about ovarian cancer subtyping:\n- There were only seven published studies on ovarian cancer subtyping, of which, three were older SVM-based approaches ([BenTaieb2015](https://link.springer.com/chapter/10.1007/978-3-319-24553-9_77), [BenTaieb2016](https://doi.org/10.4103/2153-3539.186899), [BenTaieb2017](https://doi.org/10.1016/j.media.2017.04.008)), and four were newer deep CNN approaches ([Levine2020](https://doi.org/10.1002/path.5509), [Boschman2022](https://doi.org/10.1002/path.5797), [Farahani2022](https://doi.org/10.1038/s41379-022-01146-z), [Kasture2022](https://doi.org/10.14445/22315381/IJETT-V70I3P235)).\n- All relevant studies relied on patch-level information (typically 256x256 to 512x512 pixel patches) due to the overwhelming size of whole slides. As such, multiple instance learning (MIL) methods were common.\n- The most directly relevant studies were both from UBC ([Boschman2022](https://doi.org/10.1002/path.5797), [Farahani2022](https://doi.org/10.1038/s41379-022-01146-z)), where this challenge was organised. These were the only studies in which hundreds of WSIs were classified at slide-level (not at patch-level). These studies reported AUCs well over 0.9. One of these studies has a [published code repository](https://github.com/AIMLab-UBC/ModernPath2022).\n- None of the previous studies used TMAs or had a class beside the five most common subtypes (except for non-cancer in [Kasture2022](https://doi.org/10.14445/22315381/IJETT-V70I3P235)), though there is a study on [high-grade serous vs serous borderline](https://doi.org/10.4103/jpi.jpi_76_20), and a number of studies on distinguishing tumour from benign tissue. None of the previous studies included outlier detection or uncertainty quantification. As such, this challenge is exploring uncharted territory.\n\nOutside of the review there have been some other potentially relevant papers:\n- Our paper where we found that in using whole slides for subtyping, it is typically sufficient to use only a fraction of the available information to reduce computational requirements (e.g. 10%) ([Breen2023](http://dx.doi.org/10.1117/12.2653869), [code](https://github.com/scjjb/DRAS-MIL)). \n- [Classification of Epithelial Ovarian Carcinoma Whole-Slide Pathology Images Using Deep Transfer Learning](https://openreview.net/forum?id=VXdQD8B307) An earlier short paper from UBC.\n- [Heram : Multi-Magnification Graph-Structured Whole Slide Image Representation](https://lrjconan.github.io/UBC-EECE571F-DL-Structures/assets/sample_reports_2022_W1/report_03.pdf), a multi-scale graph-based method from UBC which reports greater performance than their previous approaches, and which is currently under review.\n\nPlease do comment to add any other relevant research, and check out the previous [discussion](https://www.kaggle.com/competitions/UBC-OCEAN/discussion/445470) [posts](https://www.kaggle.com/competitions/UBC-OCEAN/discussion/445472) providing resources on the biology of ovarian cancer subtypes, previous relevant challenges, and other relevant outlier detection work", "2544010": "Adding [our newest research](https://arxiv.org/abs/2311.13956), where we found that lower magnifications actually gave better performance than higher magnifications, with performance best around 5x (where the data in this challenge starts at 20x for WSIs and 40x for TMAs). These lower magnificiations also significantly reduced the computational requirements: \nhttps://arxiv.org/abs/2311.13956 ", "2489016": "Hi Jack Breen,\nAmazing article published on Nature: \"Artificial intelligence in ovarian cancer histopathology: a systematic review\"\nhttps://www.nature.com/articles/s41698-023-00432-6" } }