Reproducibility of the GBMLGG results

[hathawayxxh]

Hi Richard,

Thank you for sharing the codes and data. I am very interested in your work. I have some concerns regarding the paper and the codes.

1. In your experiments for the pathomic fusion, the histology CNN was frozen and the SNN branch was finetuned. I am curious about why you choose to freeze the CNN branch. Have you ever tried to finetune the two branches together? Or training the two branches simultaneously?
2. In the arxiv version and the TMI version, the cindex for the histology CNN is different (0.750 vs. 0.792). Are there any difference between the settings or evaluation metrics of these two versions?
3. In your TMI paper, you said the implementation details and some results are provided in the appendix. However, I cannot find the appendix. Could you give a link for the appendix?
4. In your paper, there is a sentence "For cancer grade classification, we did not use mRNAseq expression due to missing data, lack of paired training examples, and because grade is solely determined from histopathologic appearance". If I understand correctly, it means the histological grading task is only related to the histology images rather than the omics data, so why would the pathomic fusion lead to performance gains on this task?

Looking forward to your reply. Thanks a lot for your time and attention.

Best,
Xiaohan

[Richarizardd]

Hi @hathawayxxh, thank you for your interest!

1. Histology CNN was frozen due to mitigate overfitting and training over-parameterized models with low sample-size data. In subsequent work, we have also explored training set-based network approaches for survival prediction of WSIs with multimodal fusion with genomics. In this work, the aggregation layers are trained together with Genomic SNN, but the CNN for extracting features is still frozen.

2. I believe you are looking at Version 1 of the arXiv. Please see the most recent arXiv version (Version 3).

3. The appendix and supplementary details are found in the most recent arXiv version as well. I will see to it that the TMI version is also up-to-date.

4. Your understanding is correct. Histological grading is determined via subjective interpretation of diagnostic slides via human pathologists. Though not a problem that would require multimodal data, molecular features such as IDH1 mutation and 1p/19q status are known to also correlate with histologic grades and subtypes, which is why performance improved. Though not explored in this paper, there is a lot of intrigue in the biomedical community towards understanding histology-omic correspondences and shared mutual information between these two modalities. For grade classification, one way we can think about this is that molecular features are similar in function to a "latent space" in contributing towards the phenotypic manifestation of morphological features, which then contributes towards grade prediction.

[hathawayxxh]

Thanks for your reply. My confusion is resolved.