Task Objective¶

This task extends Task 2 by integrating histopathology and transcriptomics to predict recurrence in HR-NMIBC patients. The aim is to model patient-level time-to-recurrence using both morphological and molecular data. Schematic overview of the multimodal prediction pipeline. Histopathology, RNA-seq and clinical data are encoded using pretrained networks and combined for prediction. Please note that RNA-seq data is derived from a selected tumor region within the histopathology slide.

Evaluation Metric¶

1. Model performance is evaluated using the censored concordance index (C-index). This metric measures the proportion of all comparable patient pairs where the model correctly predicts the ordering of outcomes.

2. Two patients are considered comparable if:

   • Both experienced the event (e.g., recurrence) at different times, or
   • One experienced the event, and the other was event-free, but with a longer observed follow-up time

3. A pair is not comparable if both patients experienced the event at the same time.

4. A pair is considered concordant if the patient with the higher predicted risk score has a shorter actual survival time. In other words, the model correctly orders the two patients in terms of risk.

5. The C-index ranges from:

   • 0.5 → random predictions
   • 1.0 → perfect concordance

The complete evaluation pipeline, including code for computing the censored concordance index, will be made publicly available to ensure transparency and reproducibility.

Data Details¶

Training Data¶

• 🧠 Histopathology: A single H&E-stained whole slide image (WSI) per patient, with 0.25 µm/pixel resolution at its highest resolution. Note that this WSI is either of an adjacent section of the H&E slide used for bulk RNA-seq, the same H&E slide with a punched cavity on the tissue section, or an H&E slide of another tumor of the same patient.

• 🧠 Histopathology: Binary tissue mask outlining the tissue section

• 🧬 Transcriptomics: Bulk RNA-seq data extracted from selected tumor regions, normalized using DESeq2

• 📋 Clinical Data: Same variables as Task 2.

Data versions¶

v1¶

• 126 paired multimodal training data (_HE.tif, _HE_mask.tif, _CD.json, and _RNA.json).
• Contains incorrect histopathology slides and/or tissue mask (_HE.tif, _HE_mask.tif):
  • Corrupted: 3A_024,
  • Incorrect spacings: 3A_017, 3A_031, 3A_042, 3A_050, 3A_141, 3A_143, and 3A_157
  • Scan is out of focus, resulting in failed tissue segmentation: 3A_025 (blank mask)

v2¶

• 176 paired multimodal training data (_HE.tif, _HE_mask.tif, _CD.json, and _RNA.json). Note that all the clinical data (_CD.json) files have been updated to reflect the features used in validation and test.
• These slides are fixed: 3A_024, 3A_017, 3A_031, 3A_042, 3A_050, 3A_141, 3A_143, and 3A_157. 3A_025 slide which is out of focus, however, cannot be fixed.
• Added extra materials to support model training:
  • Feature embeddings (*.pt) and
  • Coordinate of the patches (*.npy), where both are
    • extracted on the _HE.tif and _HE_mask.tif of each case,
    • using UNI at 0.25mpp resolution with 224×224 patch size, and
    • using slide2vec
• Note: The newly added 50 RNA-seq files (_RNA.json ) each begins with a prefix "3B", which indicates the source "Cohort B" as described in [1]. These Cohort B samples were sequenced using a different protocol than those from Cohort A (labeled with a prefix "3A"). Please note that no batch effect adjustment was performed on these 2 cohorts.
• UPDATE on 2025-06-08: The recently released V2 dataset contains clinical data files (_CD.json) with a missing parameter ("progression"). The clinical data files have been corrected and reuploaded to AWS, ready to be downloaded via the AWS command line below.

Download Training Data¶

Instruction (latest version)¶

1. Install AWS CLI https://docs.aws.amazon.com/cli/latest/userguide/getting-started-install.html
2. Bucket name: s3://chimera-challenge/v2/task3/
3. Command line: aws s3 sync --no-sign-request s3://chimera-challenge/v2/task3/ <destination_path>

Bucket structure (latest version)¶

v2/
  task3/
    data/ 
      task3_quality_control.csv 
      {patient_id}/
        {patient_id_CD.json}
        {patient_id_HE.tif}
        {patient_id_HE_mask.tif}
        {patient_id_RNA.json}
    features/
      coordinates/
        {patient_id_HE.npy}
        {patient_id_HE.npy}
        {patient_id_HE.npy}
      features/
        {patient_id_HE.pt}
        {patient_id_HE.pt}
        {patient_id_HE.pt}

Reference¶

1. de Jong FC, Laajala TD, Hoedemaeker RF, Jordan KR, van der Made AC, Boevé ER, van der Schoot DK, Nieuwkamer B, Janssen EA, Mahmoudi T, Boormans JL. Non–muscle-invasive bladder cancer molecular subtypes predict differential response to intravesical Bacillus Calmette-Guérin. Science translational medicine. 2023 May 24;15(697):eabn4118.