AI segmentations

I've reviewed the AI segmentations provided, and most of them look great! However, there are some cases in which the segmentation is either shifted or doesn't have the same shape as the prostate, and I wonder if it is possible to provide more accurate segmentations for these cases. Perhaps there was an issue in processing the data?

Fold 0: 10078_1000078, 10649_1000665, 10730_1000746 Fold 1: 10096_1000096, 10211_1000215, 10473_1000481, 10989_1001008 Fold 2: 10688_1000704, 10804_1000820 Fold 3: 10028_1000028, 10147_1000149, 10339_1000345, 11050_1001070, 10551_1000563 Fold 4: 10355_1000361, 11375_1001398, 11475_1001499

Additionally, there are several others in which the segmentation includes a bit of bladder or underestimates the PZ, but in general these match the prostate shape.

Best, Sara

Hi Sara,

Unfortunately, it isn't possible to provide more accurate AI segmentations without developing and retraining an all-new model for whole-gland segmentations, or manual revision of erroneous annotations. Our current whole-gland segmentation model represents our institutional best thus far (given the limited datasets available for this task), and manual annotations/revisions aren't feasible at our end (given that developing/validating whole-gland segmentations are outside the scope of this challenge). We do not recommend participants to manually handle training data correction/clean-up either, as during the Closed Testing Phase, the top 5 teams will need to retrain their models on the Private Training Set (7500-9500 additional, unseen training cases) in a fully-automated workflow using Docker containers (where manual intervention will not be possible).

Similar to the baseline AI models, our only goal for providing pseudo-annotations of the prostate gland and csPCa for all training cases (both the Public Training and Development Set, and the Private Training Set later in the challenge), is to kickstart and assist participants' model development cycle. Participants may choose to use all, some or none of these resources to assist their model training/development. Note, that baseline AI models and pseudo-annotations of the prostate gland or csPCa, will not be available during inference on any of the hidden validation or testing cases.

Hope this helps.

Hi Sara,

Thanks for reviewing the automatic prostate segmentations! I’m sure this will prove useful for future analysis and other endeavours.

As Anindo indicated, we do not plan to provide updated segmentations for these cases from our end. However, we do invite community contributions of human-expert prostate segmentations! Also, if someone would like to share their prostate segmentation algorithm and accompanying prostate segmentations, these are very welcome. To make a contribution, please check out picai_labels.

Kind regards, Joeran