Model Card for Model ID
This model is used to classify presence of pancreatic cancer.
Model Details
Input: 96x95x96x1 volume contains segmented pancreas. Use another model for segmentation (for example: TotalSegmentor or another pancreas segmented model in my repo) Voxwls must be in range [-1, 1]
Output: probability of cancer presence
0 - no cancer
1 - cancer present
Model Description
To reduce incorrect predictions pancreas should be segmented from the rest of the internal organs.
Model is a simple convolutional network with binary classification at the end.
The model was trained on custom dataset
-
50 patients CTs with cancer
-
50 patients CTs without cancer
-
Developed by: Ivan Kuchin
-
Model type: Classification
Uses
Exploration only.
Out-of-Scope Use
Model is intended for non-production use. Do not use the model as a sole source to place a diagnosis.
Bias, Risks, and Limitations
Trainig set included only 80 examples (validation set 20 examples), all samples were collected from a single CT-scanner, means model might be biased toward that scanner.
Limited number of training set samples doesn't cover all possible desease variations.
How to Get Started with the Model
Usage pipeline:
- Segment pancreas from input format using any segmentation model (TotalSegmenter or pancreas segmentation from my repository)
- Rescale to input size 96x96x96x1
- Use this model to make a prediction about cancer presence
Full inference code could be found
hereTraining Details
Trained for 1000 epochs, but early stopping finished in about 100 epochs.
Training Data
The model was trained on custom dataset
- 50 patients CTs with cancer
- 50 patients CTs without cancer
Preprocessing
Input: 96x95x96x1 volume contains segmented pancreas. Use another model for segmentation (for example: TotalSegmentor or another pancreas segmented model in my repo) Voxwls must be in range [-1, 1]
Training Hyperparameters
- Training regime: [More Information Needed]
Evaluation
20% of the full training set dedicated to a validation set
Testing Data
Due to extremely small dataset, no test set has been carved out.
Factors
Dataset has been collected from a single hospital and potentially a sole CT-scanner
Results
Final F1-score is 76% on the validation set.
Summary
Model Examination [optional]
[More Information Needed]
Environmental Impact
Carbon emissions can be estimated using the
Machine Learning Impact calculator presented in
Lacoste et al. (2019).
- Hardware Type: Tesla V100
- Hours used: 1 hout
- Cloud Provider: GCP
- Compute Region: Central
- Carbon Emitted: minute amount
Technical Specifications [optional]
Model Architecture and Objective
Convolutional classification network
Compute Infrastructure
GCP
Software
Tensorflow
