XSPATIO

XSPATIO is an explanatory computational pathology pipeline that directly links H&E morphology with spatially resolved molecular expression using ROI-aware segmentation, foundation-model feature extraction, and spatially constrained multiple-instance learning (MIL).

The framework is designed for datasets with region-level molecular ground truth, such as NanoString GeoMx DSP, enabling biologically grounded prediction and interpretation of gene and protein expression from routine histology.

Overview

XSPATIO consists of five modular stages:

1. ROI-aware segmentation (XSPATIO-SEG)
2. Patch-level feature extraction (XSPATIO-FEAT)
3. Spatially constrained MIL modeling (XSPATIO-MIL)
4. Model evaluation and cross-validation
5. Attention-based spatial visualization (XSPATIO-Heatmaps)

XSPATIO-SEG

Input Slides

ORIGINAL_ROI_FOLDER/
├── ROI_1.jpg
├── ROI_2.jpg
└── ...

Once we have the ROIs, we proceed with segmenting regions of interest using dsp coordinates available in the presets folder under XSPATIO-SEG.

python3 create_patches_fp.py \
  --source /home/ubuntu/CLAM/ORIGINAL_ROI_FOLDER \
  --save_dir SEG_patches \
  --patch_size 256 \
  --preset dsp.csv \
  --seg --patch --stitch \
  >tma_segmentation.log 2>&1 &

XSPATIO-FEAT

Once we have the segmented patches, we proceed with extracting features using the UNI model.

python3 extract_features_fp.py \
  --data_h5_dir SEG_patches \
  --data_slide_dir ORIGINAL_ROI_FOLDER \
  --csv_path SEG_patches/process_list_autogen.csv \
  --model_name uni_v1 \
  --feat_dir tma_extracted_features \
  --batch_size 512 \
  --slide_ext .jpg \
  >tma_feature_extraction.log 2>&1 &

tma_extracted_features/ ├── h5_files/ ├── pt_files/ └── feature.txt

XSPATIO-MIL

Gene or protein expression values are binarized into low- and high-expression groups using their respective thresholds across ROIs.

python3 gene_file_creater_clam.py \
  --csv_file combined_expression_tma1+2_gene_selected.csv \
  --column_name BIOMARKER_NAME \
  --h5_source_folder tma_extracted_features/h5_files \
  --pt_source_folder tma_extracted_features/pt_files \
  --low_h5_folder GENE_EXP/low_expression_genes/h5_files \
  --low_pt_folder GENE_EXP/low_expression_genes/pt_files \
  --high_h5_folder GENE_EXP/high_expression_genes/h5_files \
  --high_pt_folder GENE_EXP/high_expression_genes/pt_files

CSV Generation for MIL Training

We create CSV files with patient_id, ROI_id, and labels before training using the script below.

python3 clam_csv_generator.py

Make necessary changes in the script before running it.

low_expression_genes_folder = "GENE_EXP_CLAM/low_expression_genes"
high_expression_genes_folder = "GENE_EXP_CLAM/high_expression_genes"

output_csv_path = "GENE_EXP_FILE.csv"

The GENE_EXP_FILE.csv file should look like:

patient_id,slide_id,label
patient_0,slide_1,low_expression_genes
patient_1,slide_2,high_expression_genes

Next, we define the task in create_split_seq.py, main.py, and eval.py.

We create splits using the following command:

python3 create_splits_seq.py --task gene_exp --seed 1 --k 12

--task is the task we just defined, gene_exp

--k is the number of folds, to make splits in the dataset. 10-fold follow distribution of (80/10/10) splits for the dataset. Ours was k=12 or (76/12/12).

To run the training without interruptions run the command: tmux new -s “session”

For training, run the following command:

CUDA_VISIBLE_DEVICES=0 python3 main.py \
  --drop_out 0.25 \
  --early_stopping \
  --lr 2e-4 \
  --k 12 \
  --B 3
  --split_dir gene_exp \
  --exp_code gene_exp_100 \
  --weighted_sample \
  --bag_loss ce \
  --task gene_exp \
  --model_type clam_sb \
  --log_data \
  --data_root_dir GENE_EXP \
  --embed_dim 1024 \
  >train.txt 2>&1 &

To evaluate the script execute the command below:

CUDA_VISIBLE_DEVICES=0 python3 eval.py \
  --k 12 \
  --models_exp_code SAVED_MODEL_RESULTS \
  --save_exp_code EVAL_MODEL_RESULTS \
  --task gene_exp \
  --model_type clam_sb \
  --results_dir results \
  --data_root_dir GENE_EXP \
  --embed_dim 1024 \
  >eval.txt 2>&1 &

Attention-Based Heatmap Generation

Navigate to Heatmap/heatmaps/configs/config_template_dsp.yaml and customize the configuration template by specifying the biomarker name, corresponding trained model checkpoint, and desired output directories for heatmap generation.

CUDA_VISIBLE_DEVICES=0 python3 create_heatmaps_dsp.py \
  --config config_template_dsp.yaml