Self-Distillation Multi-Expert Ensemble Learning for Predicting Chemotherapy Efficacy of Gastric Cancer

Introduction

The Pytorch official implementation of: Self-Distillation Multi-Expert Ensemble Learning for Predicting Chemotherapy Efficacy of Gastric Cancer

Requirements

python 3.9

torch 2.0.1

torchaudio 2.0.2 

torchvision 0.15.2    

cuda 11.8

Please refer to the specific details in requirements.txt.

Running the code

1. Setting up the Python environment

   The code has been tested with Python 3.9.19. The necessary libraries can be installed using pip with the following command: #from SMEL/ pip install -r requirements.txt

2. Dataset setting

   A dataset function is required to read your own data. The SMEL model needs both image data and radiomics data. The dimension of the extracted and filtered radiomics data determines the setting of the parameter extra_dim.

3. Running the code

   You can now run the code on your dataset. You need to write a shell script file (with the extension .sh). Below is an example. You can set your parameters in the shell script file.

   Train:

   #!/bin/bash
   export MASTER_PORT=6255
   export CUDA_VISIBLE_DEVICES=0
   model_type="ResNeXt_Attention_Gating_Moe"
   logit_method='ResNeXt_Attention_Gating_Moe'
   criterion_type="ajs_uskd_mixed"
   for fold in {0..9}
       do
       next_fold=$(($fold + 1))
       #/data/wuhuixuan/code/Self_Distill_MoE/out/ResNeXt_Attention_Gating_Moe/num_expert3/ajs_uskd_mixed/1/1.0_1.0_1.0/var/No_Use_Weights/Use_criterion_total/final_output/Focal_weight_1.0/train/ResNeXt_Attention_Gating_Moe/model_best.pth.tar
       resume="/data/wuhuixuan/code/Self_Distill_MoE/out/$model_type/num_expert3/$criterion_type/${next_fold}/1.0_1.0_1.0/var/No_Use_Weights/Use_criterion_total/final_output/Focal_weight_1.0/train/$model_type/model_best.pth.tar"
       python main.py --fold $fold --model_type "$model_type" --batch-size 8 --epoch 150 --logit_method "$logit_method" --criterion_type "$criterion_type" --logit_alpha 0.5 --logit_beta 0.5 --logit_gamma 0.5 --weights 1 1 1 --use_criterion_total True --extra_dim 22 --diversity_metric "var" --extra_input True #--evaluate --resume $resume --display_experts True
   done
   
   

   Test:

   #!/bin/bash
   export MASTER_PORT=6255
   export CUDA_VISIBLE_DEVICES=0
   model_type="ResNeXt_Attention_Gating_Moe"
   logit_method='ResNeXt_Attention_Gating_Moe'
   criterion_type="ajs_uskd_mixed"
   for fold in {0..9}
      do
      next_fold=$(($fold + 1))
      #/data/wuhuixuan/code/Self_Distill_MoE/out/ResNeXt_Attention_Gating_Moe/num_expert3/ajs_uskd_mixed/1/1.0_1.0_1.0/var/No_Use_Weights/Use_criterion_total/final_output/Focal_weight_1.0/train/ResNeXt_Attention_Gating_Moe/model_best.pth.tar
      resume="/data/wuhuixuan/code/Self_Distill_MoE/out/$model_type/num_expert3/$criterion_type/${next_fold}/1.0_1.0_1.0/var/No_Use_Weights/Use_criterion_total/final_output/Focal_weight_1.0/train/$model_type/model_best.pth.tar"
      python main.py --fold $fold --model_type "$model_type" --batch-size 8 --epoch 150 --logit_method "$logit_method" --criterion_type "$criterion_type" --logit_alpha 0.5 --logit_beta 0.5 --logit_gamma 0.5 --weights 1 1 1 --use_criterion_total True --extra_dim 22 --diversity_metric "var" --extra_input True --evaluate --resume $resume --display_experts True
   done
   
   

   Note: If you do not need ten-fold cross-validation, you can set the fold to any number or leave it unset. You should remove the for loop statement in the .sh file. Note: --evaluate sets the evaluation mode; otherwise, it is training mode. --resume is the path to the model parameters file, and if you need to output images from each expert model, set --display_experts to True.