File size: 2,032 Bytes
232c308 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 | ---
license: mit
task_categories:
- tabular-classification
- image-classification
- feature-extraction
tags:
- medical-imaging
- glioma
- multi-task-learning
- survival-analysis
- cmtlnet
---
# CMTLNet: Cooperative Multi-Task Learning for Glioma Subtyping & Survival Analysis
This repository contains the checkpoints, trained models, logs, results, and source code for the **CMTLNet** framework.
## 🏆 Step 1: 5-Fold Cross-Validation CFE Performance
Stage 1 trains the **Common Feature Extractor (CFE)** across all 5 folds (60 epochs each) on three cooperative clinical tasks:
1. **1p/19q Co-deletion** (INT vs CODEL)
2. **IDH Mutation** (Wildtype vs Mutant)
3. **Tumor Grade (LHG)** (LGG Grade 2/3 vs HGG Grade 4)
### Cross-Validation Summary:
| Fold | Epochs | Best Test F1 (Macro) | Best Accuracy | Best Epoch | Checkpoint Path in Repo |
| :---: | :---: | :---: | :---: | :---: | :--- |
| **Fold 0** | 60 | **78.04%** | 91.42% | Epoch 7 | `best_models/fixedone/Fold_0/CFE_F1_78.04_step_3088.pkl` |
| **Fold 1** | 60 | **73.22%** | 93.00% | Epoch 13 | `best_models/fixedone/Fold_1/CFE_F1_73.22_step_5432.pkl` |
| **Fold 2** | 60 | **77.30%** | 89.67% | Epoch 7 | `best_models/fixedone/Fold_2/CFE_F1_77.30_step_3112.pkl` |
| **Fold 3** | 60 | **73.91%** | 85.33% | Epoch 4 | `best_models/fixedone/Fold_3/CFE_F1_73.91_step_1945.pkl` |
| **Fold 4** | 60 | **78.47%** | 88.33% | Epoch 6 | `best_models/fixedone/Fold_4/CFE_F1_78.47_step_2744.pkl` |
**Mean 5-Fold Cross-Validation Test F1:** **76.19%**
---
## 📁 Repository Structure
- `best_models/fixedone/Fold_{0..4}/`: Best performing CFE model weights for each fold.
- `checkpoints/fixedone/`: Intermediate training checkpoints saved every 5 epochs and final epoch 59.
- `results/CFE/Fold_{0..4}/`: Complete epoch-by-epoch CSV evaluation metrics (`log_mean.csv`, `log_1p19q.csv`, `log_IDH.csv`, `log_LHG.csv`).
- `logs/`: Complete runtime logs (`step1_nohup.log`).
- `code/`: Complete source code, models, dataset split pickles, and config files for Step 1 & Step 2.
|