Please enable JavaScript.
Coggle requires JavaScript to display documents.
Nature-comm revision - Coggle Diagram
Nature-comm revision
Code part
-
32 GB is recommened but not required. Instead, more host memory size is required to train huge images.
-
In the intiialization_data.py, it is said that slides without labels are accepted and will be assigned as "-1".
-
Description part
-
Minor
-
-
:check: In Figure 2, 3-way classifier or 2-way classifier
:warning: slide size distribution. Is 20,000x20,000 original size or resized size. Clarify these.
-
:green_cross: 10k slide is not needed, and not trainable due to host memory constraint. (no longer a problem since the experiments are expanded.)
-
:check: Why the heatmap of MIL so smooth? Because of interpolation. (need new visualization results)
-
:check: Page 3: “The main idea of MIL on slide-level cancer classification is that if the patches with highest scores (the most possible K patches) on the slide were identified as carcinoma, the slide should be classified into cancer, and vice versa.”
:check: “Moreover, recent studies show that even state-of-the-art weak supervision methods still cannot attain the average performance of strong supervision methods in most computer vision fields such as object detection, semantic segmentation, and instance segmentation tasks.”
You should cite appropriate papers when making such statements
-
-
Clarify that our method does not require powerful GPU, but more host memory
-
Experiment part
-
-
Abalation study
-
:check: 2-way classifier (adeno vs others, squamous vs others) (2 tests)
:check: 2x, 10x (5 tests) :green_cross:
-
-
-
-
Related work part
Detection of prostate cancer in whole-slide images through end-to-end training with image-level labels
-
-
-