pancreas_cancer_classification

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:

1. Segment pancreas from input format using any segmentation model (TotalSegmenter or pancreas segmentation from my repository)
2. Rescale to input size 96x96x96x1
3. Use this model to make a prediction about cancer presence

Full inference code could be found here

Training 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] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->

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]

<!-- Relevant interpretability work for the model goes here -->

[More Information Needed]

Environmental Impact

<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->

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