![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F21013672%2Fc0f60caebf0af22f9cc5e38a7a8609fb%2FScreenshot%202025-12-02%20at%206.27.29PM.png?generation=1764734806047530&alt=media\")
Hi, here is a brief overview on the competetion metric.\nHope this is helpful for y'all.
\nI will refer to the predicted mask as P and ground truth mask as G.\nReasons why standard DiceLoss will not suffice
\nTo solve these problems, the metric used to evaluate our solutions is as given below:\nScore = 0.30 × TopoScore + 0.35 × SurfaceDice@τ + 0.35 × VOI_score
\nτ here is the spatial tolerance indicating spatial transformation is acceptable as long as the geometry is predicted correctly.\nCalculation of the metric goes as such:
\n
This metric measures the information loss when you map the connected components of P and G.\ncomprises of the sum of the split VOI and merge VOI where conditional entropy are calculated and the information gain is subtracted.\nSince this metric is boundless the VOI_score is calcualted as given bellow.\nVOI_score here directly peanlises over and under segmentation on the surface level.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F21013672%2F520cef18b65c7efbeaf37520822b4cea%2FScreenshot%202025-12-03%20at%209.37.24AM.png?generation=1764734886459503&alt=media\")
the topological matching is done through the toposcore.\nToposcore is the weighted average of the topological f1 scores done along each dimension.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F21013672%2F3b3c68674eb18ac53b426a136e724378%2FScreenshot%202025-12-03%20at%209.37.46AM.png?generation=1764734967208693&alt=media\")
The direct effect on mis segmentation is given in the competition overview, do check it out!!
", "messages": [ { "id": 3361149, "postDate": "2025-12-03T04:09:38.670Z", "content": "Hi, here is a brief overview on the competetion metric.\nHope this is helpful for y'all.
\nI will refer to the predicted mask as P and ground truth mask as G.\nReasons why standard DiceLoss will not suffice
\nTo solve these problems, the metric used to evaluate our solutions is as given below:\nScore = 0.30 × TopoScore + 0.35 × SurfaceDice@τ + 0.35 × VOI_score
\nτ here is the spatial tolerance indicating spatial transformation is acceptable as long as the geometry is predicted correctly.\nCalculation of the metric goes as such:
\n
This metric measures the information loss when you map the connected components of P and G.\ncomprises of the sum of the split VOI and merge VOI where conditional entropy are calculated and the information gain is subtracted.\nSince this metric is boundless the VOI_score is calcualted as given bellow.\nVOI_score here directly peanlises over and under segmentation on the surface level.
\n
the topological matching is done through the toposcore.\nToposcore is the weighted average of the topological f1 scores done along each dimension.
\n
The direct effect on mis segmentation is given in the competition overview, do check it out!!
", "rawMarkdown": "Hi, here is a brief overview on the competetion metric.\nHope this is helpful for y'all.\n\nI will refer to the predicted mask as *P* and ground truth mask as *G*.\n**Reasons why standard DiceLoss will not suffice**\n1. Standard Dice Loss gives equal weightage to all voxels and hence penalises missclassification equally. However this isnt true as missclassification of certain voxels make the virtual unwrapping of scrolls impossible.\n2. Inability of the model to detect air between 2 layers makes a prediuction of a single thick layer causing a merger acorss adjecent wraps.\n3. Over segmentation leads to splitting across the same wrap.\n\nTo solve these problems, the metric used to evaluate our solutions is as given below:\n**Score = 0.30 × TopoScore + 0.35 × SurfaceDice@τ + 0.35 × VOI_score**\n\n## SurfaceDice@τ\nτ here is the spatial tolerance indicating spatial transformation is acceptable as long as the geometry is predicted correctly.\nCalculation of the metric goes as such:\n1. isolation of surface on both *P* and *G*.\n2. Compution of min dist for all points p belonging to *Sp* to *Sg* and vice versa.\n3. aggregation of number of points whose min dist <= τ.\n4. Averaging the matched fractions on both sides\n\n\n\n\n##Variation of information\nThis metric measures the information loss when you map the connected components of *P* and *G*.\ncomprises of the sum of the split VOI and merge VOI where conditional entropy are calculated and the information gain is subtracted.\nSince this metric is boundless the VOI_score is calcualted as given bellow.\nVOI_score here directly peanlises over and under segmentation on the surface level.\n\n\n\n##TopoScore\nthe topological matching is done through the toposcore.\nToposcore is the weighted average of the topological f1 scores done along each dimension.\n\n\n\nThe direct effect on mis segmentation is given in the competition overview, do check it out!!\n", "votes": 12 }, { "id": 3362268, "postDate": "2025-12-04T12:05:15.740Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3362268, "author_name": "", "author_url": "", "post_date": "2025-12-04T12:05:15.740000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3361149": "Hi, here is a brief overview on the competetion metric.\nHope this is helpful for y'all.\n\nI will refer to the predicted mask as *P* and ground truth mask as *G*.\n**Reasons why standard DiceLoss will not suffice**\n1. Standard Dice Loss gives equal weightage to all voxels and hence penalises missclassification equally. However this isnt true as missclassification of certain voxels make the virtual unwrapping of scrolls impossible.\n2. Inability of the model to detect air between 2 layers makes a prediuction of a single thick layer causing a merger acorss adjecent wraps.\n3. Over segmentation leads to splitting across the same wrap.\n\nTo solve these problems, the metric used to evaluate our solutions is as given below:\n**Score = 0.30 × TopoScore + 0.35 × SurfaceDice@τ + 0.35 × VOI_score**\n\n## SurfaceDice@τ\nτ here is the spatial tolerance indicating spatial transformation is acceptable as long as the geometry is predicted correctly.\nCalculation of the metric goes as such:\n1. isolation of surface on both *P* and *G*.\n2. Compution of min dist for all points p belonging to *Sp* to *Sg* and vice versa.\n3. aggregation of number of points whose min dist <= τ.\n4. Averaging the matched fractions on both sides\n\n\n\n\n##Variation of information\nThis metric measures the information loss when you map the connected components of *P* and *G*.\ncomprises of the sum of the split VOI and merge VOI where conditional entropy are calculated and the information gain is subtracted.\nSince this metric is boundless the VOI_score is calcualted as given bellow.\nVOI_score here directly peanlises over and under segmentation on the surface level.\n\n\n\n##TopoScore\nthe topological matching is done through the toposcore.\nToposcore is the weighted average of the topological f1 scores done along each dimension.\n\n\n\nThe direct effect on mis segmentation is given in the competition overview, do check it out!!\n", "3362268": "" } }