This repository contains the work done by the VCMI team for the CHIMERA Challenge (Combining HIstology, Medical Imaging (Radiology), and molEcular Data for Medical pRognosis and diAgnosis).
When cloning remember to git submodule update --init to initialize the submodules! (Official baseline, external tools like Trident, etc.)
The CHIMERA Challenge is an international AI competition aiming to integrate histology, radiology, and molecular (transcriptomic) data for better cancer prognosis and diagnosis.
By benchmarking multimodal models on prostate cancer and non-invasive bladder cancer (NIBC) datasets, the challenge seeks to overcome barriers in precision medicine and multimodal data fusion.
👉 More details can be found on the official challenge page.
The VCMI team ranked 5th among 22 international teams in the challenge.
| Rank | Team Name | C-Index |
|---|---|---|
| #1 | TIA-Pegasus | 0.7402 |
| #2 | WL | 0.7294 |
| #3 | SMILE | 0.7280 |
| #4 | IUCompPath | 0.7197 |
| #5 | VCMI | 0.7153 |
| #6 | OHSU-Cedar | 0.6885 |
| #7 | CADGEN_BIIT | 0.5000 |
| ... | ... | ... |
The CHIMERA dataset is composed of 95 patients with 27 showing recurrence and 68 being censored. Each patient has 3 modalities: Histopathology (WSIs), Imaging (mpMRI), and Clinical Data.
Performance on the validation set across 10 folds (trained for 100 epochs). Values are the mean ± std of the best C-index per fold. The best result is in bold, and the second best is underlined.
Filtered version = excludes the pathology-derived values from the clinical data
| Modality | C-index (mean ± std) |
|---|---|
| C | 0.8465 ± 0.0518 |
| C (filtered) | 0.7124 ± 0.1113 |
| M | 0.7473 ± 0.1060 |
| P | 0.8224 ± 0.0746 |
| C + M | 0.8526 ± 0.0513 |
| C (filtered) + M | 0.7153 ± 0.0830 |
| C + P | 0.8902 ± 0.0601 |
| C (filtered) + P | 0.8452 ± 0.0565 |
| M + P | 0.8374 ± 0.0533 |
| C + M + P | 0.8786 ± 0.0466 |
| C (filtered) + M + P | 0.8555 ± 0.0713 |
This repository is licensed under the MIT License.