ccRCC is the most common type of kidney cancer, accounting for up to 80% of all renal cell carcinoma cases in adults. Estimating the prognosis is critical for patient management, but it is still a very challenging task. Ongoing research on this topic has led to the creation of two public studies: CPTAC-CCRCC and TCGA-KIRC, from which we curated MMIST-ccRCC.
| Modality | Balanced Accuracy (%) |
|---|---|
| CT (MIL) | 61.59 |
| MRI (MIL) | 50.00 |
| WSI (MIL) | 53.42 |
| ClinGen (MIL) | 73.16 |
Results are organized into WSI-specific prediction and Patient-specific aggregation strategies for survival prediction at 12 months using WSI-only data.
| Model | Balanced Accuracy (%) |
|---|---|
| GCN | 0.6873 ± 0.0061 |
| GAT (1 head) | 0.6917 ± 0.0118 |
| GAT (2 heads) | 0.6866 ± 0.0181 |
| GraphSAGE FB (mean) | 0.7003 ± 0.0089 |
| GraphSAGE FB (max) | 0.6998 ± 0.0135 |
| Aggregation | Model | Bacc ± Std (%) |
|---|---|---|
| Majority Voting | GCN | 67.95 ± 2.08 |
| GAT 1 head | 68.01 ± 3.36 | |
| GAT 2 head | 68.56 ± 2.42 | |
| GraphSAGE FB (mean) | 70.78 ± 4.46 | |
| GraphSAGE FB (max) | 69.38 ± 2.98 | |
| One Dominance | GCN | 70.23 ± 2.55 |
| GAT 1 head | 69.18 ± 3.18 | |
| GAT 2 head | 69.70 ± 2.61 | |
| GraphSAGE FB (mean) | 71.04 ± 2.12 | |
| GraphSAGE FB (max) | 69.49 ± 5.27 | |
| MIL WSI | GCN | 75.05 ± 0.18 |
| GAT 1 head | 72.41 ± 2.11 | |
| GAT 2 head | 76.42 ± 1.62 | |
| GraphSAGE FB (mean) | 73.46 ± 2.92 | |
| GraphSAGE FB (max) | 71.35 ± 2.13 |
| Experiment | CT | MRI | WSI | ClinGen | All Patients |
|---|---|---|---|---|---|
| Late Fusion WS | 65.00 | 73.33 | 73.16 | 73.16 | 73.15 |
| Late Fusion LW | 73.85 | 61.66 | 69.36 | 69.36 | 69.35 |
| Early Fusion Mean | 88.18 | 73.33 | 83.23 | 83.23 | 82.32 |
| Early Fusion Cat | 77.50 | 85.00 | 78.45 | 78.45 | 78.45 |
| Early Fusion Mean (W/ Reconstruction) | 90.91 | 61.67 | 84.70 | 84.70 | 84.70 |
| Early Fusion Cat (W/ Reconstruction) | 85.45 | 56.67 | 83.27 | 83.27 | 83.27 |
If you want more information about this benchmark feel free to read our papers:
Tiago Mota, Maria Rita Fonseca Verdelho, Diogo José Pereira Araújo, Alceu Bissoto, Carlos Santiago, Catarina Barata, MMIST-ccRCC: A Real-World Medical Dataset for the Development of Multi-Modal Systems, Data Curation and Augmentation in Enhancing Medical Imaging Applications Workshop (archival) @CVPR 2024.
Maria Rita Fonseca Verdelho, Alexandre Bernardino, Catarina Barata, MIL vs. Aggregation: Evaluating Patient-Level Survival Prediction Strategies Using Graph-Based Learning
Cite the paper where the dataset was first published and acknowledge us: "The results shown in this work used datasets collected from MMIST: https://Multi-Modal-IST.github.io/"