ct-counterfactuals 0.0.4

Models and methods to generate counterfactuals for computed tomography scans

CT Counterfactuals

The code and models here were used in the Paper 📄 Merlin: A Vision Language Foundation Model for 3D Computed Tomography .

Classifiers

A 1692 target classifier predicting phenotypes from CT scans

import ct_counterfactuals as ct_cf
model = ct_cf.classifiers.phecode.PheCodeClassifier()
x = torch.ones([1, 1, 224, 224, 174])
out = model(x)
out.shape # [1, 1692]

A lung segmentation model from CT slices

import ct_counterfactuals as ct_cf
model = ct_cf.classifiers.lungmask.LungMaskSegmenter()
x = torch.ones([1, 1, 224, 224, 174])
out = model(x)
out.shape # [1, 3, 224, 224, 1]

# Channels
# 0 = No lung
# 1 = Right lung
# 2 = Left lung

Autoencoders

A VQ-GAN autoencoder trained on CT slices

import ct_counterfactuals as ct_cf
ae = ct_cf.ae.VQGAN(weights='2023-12-25T10-26-40_ct2_vqgan256_sddd')
x = torch.ones([1, 1, 224, 224])
out = ae(x)
out.shape # [1, 1, 224, 224]

Utility code is provided to encode 3D volumes

import ct_counterfactuals as ct_cf
ae = ct_cf.ae.VQGAN(weights='2023-12-25T10-26-40_ct2_vqgan256_sddd')

slice_ae = SliceAEFull(ae, 45, 55) # range specified is where gradients can propigate
x = torch.ones([1, 1, 224, 224, 174])
out = ae(x)
out.shape # [1, 1, 224, 224, 174]

Example CF explainations of the classifier

Effusion (fluid in lungs)	Splenomegaly (enlarged spleen)