detection maps

Hello, I would like to inquire about the specifics of how detection maps are defined or obtained. As you know, typically, the output of a model is post-processed to represent integers ranging from 0 to num_classes. However, if the goal is to have the model's output processed into detection maps where each element falls within the range of 0-1, I'm wondering how this is achieved, especially concerning the channel dimension. Is it possible to simply sum up the channels and then remove the channel dimension? I hope to receive your guidance on this matter. Thank you！！！

Hi Hongkun Sun,

The detection map is defined as a volume (with the same dimensions and resolution as the CECT image) where for a the predicted lesion, all voxels must comprise a single floating point value between 0-1, representing that lesion’s likelihood of harboring PDAC. The PANORAMA baseline method describes their method to obtain detection maps from the softmax predictions in the repository's README file:

"The final detection map is obtained by ensembling the best checkpoints for each fold. The nnU-Net output probability map is then converted to a detection map using the GetFullSizDetectionMap function that applies the extract_lesion_candidates method from report_guided_annotation [3]. For more information about the lesion extraction process refer to the documentation in the original repository."

The reference [3] is publicly available, and for more information on the detection map formulation, you can check out its supplementary materials. Specifically, on page 2, section "Extraction of Lesion Candidates":

"From each softmax volume (a voxel-level confidence map) we create distinct lesion candidates, as illustrated in Figure S2. Specifically, we define a lesion candidate by starting at the most confident voxel and construct a region of connected voxels (in 3D) with at least 40% of the peak’s confidence. We set the lesion candidate confidence to the maximum confidence within the lesion candidate region. Then, the region is set to zero in the softmax volume, and the process is repeated until 5 lesions are extracted, or no candidates remain. Small candidates of 10 or fewer voxels (≤ 0.009 cm3) are discarded. When the highest remaining confidence is less than 0.01, the extraction is terminated. This extraction method ensures low-confidence lesions can be extracted from the softmax volume (as shown for Lesion candidate 3 in Figure S2), without accidentally merging close high-confidence lesion candidates (as shown for Lesion candidates 1 and 2 in Figure S2)"

For a visual example, check out the supplementary materials above.

Hi Hongkun Sun,

The description provided by Joeran is accurate, if you need any further clarification let me know. I would only like to emphasize that for obtaining the final PDAC detection map we only consider channel 1 of the nnU-Net predicted output, which is the channel concerning the tumor prediction. The other channels for the structures segmentations are only used to post-process the detection map as shown in the PostProcessing function.

Thank you very much for your responses, Joeran Bosma and Natália Alves. I have some further questions. First, as I understand it, "The nnU-Net output probability map" and "each softmax volume (a voxel-level confidence map)" mean the same thing; they are the results of the model's output after being processed by the softmax function. Is my understanding correct? Also, does the softmax result mentioned above only contain information from channel 1 and is unrelated to other channels?

Yes that is correct. The nnU-Net outputs detection maps (or probability maps, or softmax volumes) for every channel, but in our base line algorithm we only take into account channel 1 which is the tumor channel.