A demo implementation for reproducing the results of Whole Slide Image (WSI) search using Conditioned Deep Sparse Fisher Vector (C-Deep-SFV) and Conditioned Deep Binary Fisher Vector (C-Deep-BFV) on the diagnostic slides from The Cancer Genomic Atlas (TCGA) repository presented in the following paper.
“Learning Binary and Sparse Permutation-Invariant Representations for Fast and Memory Efficient Whole Slide Image Search”.
To reduce size of dataset, we already extracted the features of WSIs patches. Please download the dataset here, unzip the dataset, set the current working directory to a folder that contains “PipelineConfig_cluster_5_tn_700_tp1_45_tp2_45”, "gdc_data.csv", “WSI_search_C_Deep_SFV.py”, and “WSI_search_C_Deep_BFV.py” files and then run the WSI_search_C_Deep_SFV.py and WSI_search_C_Deep_BFV.py to reproduce the WSI search results for C-Deep-SFV and C-Deep-BFV embeddings. These results will be for the C-Deep-SFV and C-Deep-BFV columns of Table 1 in the paper. Due the fact we are employing a variational sutoencoder as the deep generative model for WSI representation learning, there might be small variations in the results compared with paper.
In order to run demo locally, one tested working configuration is to create an anaconda environment and follow these steps on Anaconda Prompt:
- conda create -n compact_deep_FV python=3.9.7
- conda activate compact_deep_FV
- numpy==1.21.5
- pandas==1.1.5
- scikit_learn==1.1.1
- scipy==1.7.3
- tensorflow==2.6.0
This procedure has been tested on a local machine with Windows 10 (64 bit) with the following specs:
- RAM: 64.0 GB RAM
- CPU: Intel(R) Core(TM) i9-9900X 3.50 GHz
- GPU: NVIDIA GeForce RTX 2080 SUPER GPU