Lung Cancer BOMI2 Dataset

This repository contains patient-level metadata, tissue sample information, and cross-validation splits for a cohort of 298 patients with non-small cell lung cancer (NSCLC).
Use this dataset to validate and compare with our different methods

It is intended for machine learning and computational pathology research, in particular for:

• Binary survival prediction (high vs. low survival)
• Binary subtype classification
  • adenocarcinoma vs. squamous cell carcinoma + others
• Binary subtype classification (restricted)
  • adenocarcinoma vs. squamous cell carcinoma (excluding “others”)

---

Cohort Overview

• Patients: 298
• Tissue samples: 500 cores (multiple cores per patient are possible)
• Stains: CD4, CD8, FoxP3, CD20, PanCK, and DAPI
• Features:
  • Extracted cell coordinates, marker intensities, cell phenotypes, and nuclear morphology features (see link below)
  • Clinical data: survival times, censoring, tumor stage, age, sex, histological subtype

---

Repository Structure

lung_cancer_BOMI2_dataset/
├── binary_survival_prediction/
│ ├── 100foldcrossvalrepeat/
│ │ ├── split_0_train.csv
│ │ ├── split_0_val.csv
│ │ ├── split_0_test.csv
│ │ └── ...
│ ├── 10foldcrossvalrepeat/
│ │ ├── split_0_train.csv
│ │ └── ...
│ ├── static_split/
│ │ ├── train.csv
│ │ ├── val.csv
│ │ ├── train_val.csv
│ │ └── test.csv
│ ├── Clinical_data_with_labels.csv
│ └── samples_labels.csv
│
├── binary_subtype_prediction/
│ └── (same structure as above)
│
├── binary_subtype_prediction_ACvsSqCC/
│ └── (same structure as above)
│
└── README.md

---

File Descriptions

Cross-validation splits

• 100foldcrossvalrepeat/
  10×10 repeated cross-validation. Each split_k_* file contains patient IDs for train, validation, or test for that fold.
• 10foldcrossvalrepeat/
  Standard 10-fold cross-validation (single repetition).
• static_split/
  Fixed data split for reproducibility:
  • train.csv
  • val.csv
  • train_val.csv
  • test.csv

Metadata

• Clinical_data_with_labels.csv
  Patient-level clinical metadata combined with labels for the prediction task.
• samples_labels.csv
  Tissue-sample-level data with patient IDs and labels.

---

Prediction Tasks

1. Binary survival prediction
   Label: Above median vs. below median overall survival.

2. Binary subtype classification (all patients)
   Label: adenocarcinoma vs. squamous cell carcinoma and “others”.

3. Binary subtype classification (restricted)
   Label: adenocarcinoma vs. squamous cell carcinoma.
   Patients with “others” subtype are excluded.

---

Cell-level Features and Raw Data

Cell-level features (coordinates, marker intensities, phenotypes, and nuclear morphology) can be accessed at:

[link to zenodo]

These features can be linked to the patient and sample IDs provided in this repository.

create_samples.py

From raw_data/Clinical_data_max_20190329.csv (clinical params), and raw_data/TIL_tissue_seg_data.txt (core level info), and raw_data/BOMI_TIL_merged_cell_seg_data_feats_1.csv (cell level info) and raw_data/id_align.csv (patient <-> core mapping), generate, in binary_survival_prediction, Clinical_data_with_labels.csv (selected clinical params, and label = 'days>=median_days'), and samples_labels.csv (tumor and stroma areas for each core, mapped to patient),

split_dataset.py

From binary_survival_prediction/Clinical_data_with_labels.csv, generate, in binary_survival_prediction, 10foldcrossval/ and 100foldcrossvalrepeat/

generate_dataset_profile_report.py

From binary_survival_prediction/Clinical_data_with_labels.csv, generate report.html with basic stats about the dataset.