Doubt on slide or patient level prediction

[Rukhmini]

Hello,

Did you apply the survival prediction on slide level or patient level? If it is on patient level can you please let me know how did you convert the slide level prediction to patient level prediction?

[LiangJunhao-THU]

Hi, Rukhmini, thank you for your question. In this study we selected the largest tumour fraction one as the patient’s representative WSI, which means we used the selected WSI score as the patient score. This phenomenon is quite common in similar studies, primarily due to the uncertainty in the number of patient slides available in the dataset. For instance, in TCGA, most patients only have one corresponding whole-slide image (WSI), whereas hospitals may collect patient data with varying numbers of WSIs, ranging from one to several dozen. We extensively discuss this issue in the "Robustness of macro mode indicators" section of the paper.

[Rukhmini]

Thank you Junhao for your response. That means you trained the model with all the slides and get slide level risk score first and then selected the risk score for the slide that has largest tumor fraction as patient level score. Also, do you have the code for the random selection as you mentioned in the "Robustness of macro mode indicators" section of the paper.

[LiangJunhao-THU]

Thank you for your interests. It's not entirely as you described. Just as you initially raised, we did consider this issue during the experimental design phase. After completing the tissue segmentation in the initial step, we were able to calculate the tumor's proportion on each WSI. Before conducting prognostic experiments, we predetermined to select the WSI with the largest tumor area as the representative score for each patient. Therefore, we didn't train the model using all WSIs; rather, we trained it using one WSI per patient selected according to the predefined rule. This approach minimally affects training on TCGA data, where each patient typically has only one WSI. After training the prognostic model, we became curious about its robustness and the issue of multiple slices per patient. Hence, we evaluated the trained model's performance by inferring all WSIs to explore its robustness.

[Rukhmini]

Thank you for the clarification.