File size: 2,918 Bytes
f91c847 da0fcc1 f91c847 | 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 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 | ---
language: en
license: mit
library_name: pytorch
tags:
- medical-imaging
- mri
- 3d-resnet
- monai
- oncology
- breast-cancer
datasets:
- odelia
metrics:
- roc_auc
pipeline_tag: image-classification
---
# 3D ResNet-18 for ODELIA MRI Classification
This repository contains the weights for a 3D ResNet-18 model trained for the ODELIA dataset (multi-parametric MRI). The model was developed to classify medical images into 3 distinct classes using 5-fold cross-validation.
## Model Description
- **Developed by:** THOUAN Simon
- **Model type:** 3D Convolutional Neural Network (ResNet-18 architecture)
- **Framework:** PyTorch & MONAI
- **Task:** Multi-class 3D Image Classification (3 classes)
- **Input:** 5 MRI sequences (Pre, Post_1, Post_2, Sub_1, T2) concatenated as channels.
- **Input Size:** (128, 128, 64)
## Architecture Details
The model is a 3D adaptation of the ResNet architecture provided by the **MONAI** library:
- **Blocks:** Basic block
- **Layers:** [2, 2, 2, 2] (Equivalent to ResNet-18)
- **In-planes:** [64, 128, 256, 512]
- **Input Channels:** 5
- **Spatial Dimensions:** 3D
## Training Procedure
The model was trained on a high-performance computing cluster (IDUN) using the following configuration:
### Hyperparameters
| Parameter | Value |
| :--- | :--- |
| **Optimizer** | Adam |
| **Learning Rate** | 1e-4 |
| **Loss Function** | CrossEntropyLoss |
| **Batch Size** | 4 (Training) / 2 (Validation) |
| **Epochs** | 50 |
| **Validation** | Every 2 epochs |
### Cross-Validation Strategy
The dataset was split into 5 folds (A, B, C, D, E) to ensure robustness.
- **Fold 0:** Val = A, Train = B+C+D+E
- **Fold 1:** Val = B, Train = A+C+D+E
- *(Continuing for all 5 folds)*
### Preprocessing (MONAI Transforms)
- **Resizing:** All volumes resized to $128 \times 128 \times 64$.
- **Normalization:** Intensity scaling for each sequence.
- **Concatenation:** The 5 MRI sequences are stacked into a single 5-channel tensor.
## Evaluation Results
The primary metric used is **ROC AUC** (Area Under the Receiver Operating Analytic Curve).
Detailed performance graphs and confusion matrices for the ensemble can be found in the associated [GitHub Repository](https://github.com/THOUAN-Simon/ODELIA_Challenge_CV_DL)).
## How to Load the Model
```python
import torch
from monai.networks.nets import ResNet
# Initialize architecture
model = ResNet(
block="basic",
layers=[2, 2, 2, 2],
block_inplanes=[64, 128, 256, 512],
n_input_channels=5,
num_classes=3,
spatial_dims=3
)
# Load weights
state_dict = torch.load("weights/best_resnet_odelia_fold0.pth", map_location="cpu")
model.load_state_dict(state_dict)
model.eval()
```
## Limitations & Ethical Considerations
This model is for research purposes only. It was trained on the ODELIA proprietary dataset. Predictions should not be used for clinical diagnosis without professional medical supervision.
|