A method for detecting breast cancer from thermal images using bilateral symmetry and learned texture features. Featured in ISBI 2025: https://biomedicalimaging.org/2025/full-program/
We train an encoder to disentangle texture and structure attributes from the images, following the principles of the Swapping Autoencoder by Park et al. (https://arxiv.org/abs/2007.00653) and Lung-Swapping Autoencoder by Zhou et al. (https://arxiv.org/abs/2201.07344).
We compare the texture attributes between left and right breasts with a simple absolute difference between the texture vectors. A classifier is then trained over these differences, thus making it able to efficiently detect normal cases from cancerous ones.
The dataset used in this project can be downloaded here: https://www.kaggle.com/datasets/tiennelescarbeault/dmr-ir-breast-cancer-detection-for-tada-sae/data
It comprises the formatted breast segmentation masks for both normal and anomalous cases and train-test splits. The original dataset can be found here: http://visual.ic.uff.br/en/proeng/thiagoelias/
It comprises of thermal images from 57 patients, in which 36 were diagnosed with breast cancer. In a thermal image, this is characterised by a temperature increase over the affected side. In this version, left and right breasts were extracted separately, which is useful to compare their asymmetries, thus indicating possible breast cancer.
To reproduce the experiments found in the paper, 3 files are provided: run_dmrir_baselines.py, run_dmrir_ablation.py and run_dmrir_tadasae.py.
The file run_dmrir_baselines.py contains the code to reproduce the baseline experiments, which are supervised classification using ResNet-18, InceptionV3 and a simple CNN. The parameters can be adjusted in the command line or in the configs/supervised_dmrir.yaml file.
Example: python run_dmrir_baselines.py --experiment resnet_18
The file run_tadasae_ablation.py contains the code to reproduce the ablation experiments, which are the experiments to evaluate the importance of each component of the TADA-SAE method. The parameters can be adjusted in the command line or in the configs/ae_dmrir.yaml or configs/tadasae_dmrir.yaml files, depending on the chosen ablation experiment. Three (3) ablation experiments are provided:
-
ae_full_im: Simple autoencoder to extract image features from the whole image, followed by a classification. Therefore, no texture disentanglement or left-right symmetry is used.
-
ae_left_right: Simple autoencoder to extract image features from the left and right images, followed by a classification. Therefore, no texture disentanglement is used but does left-right comparisons.
-
lsae_full_im: A swapping autoencoder based on LSAE (Zhou et al., 2022) to extract image features from the whole image, followed by a classification. Texture disentanglement is performed but no left-right comparison is conducted.
Example: python run_dmrir_ablation.py --experiment ae_full_im
The file run_dmrir_tadasae.py contains the code to reproduce the TADA-SAE method. The parameters can be adjusted in the command line or in the configs/tadasae_dmrir.yaml file. The TADA-SAE method is a combination of the three components: texture disentanglement, left-right symmetry and classification. It is based on the LSAE (Zhou et al., 2022) autoencoder, which is used to extract image features from the left and right images, followed by a classification. For this model, weights are provided through this link:
https://www.kaggle.com/models/tiennelescarbeault/tada-sae-swapping-autoencoder-dmr-ir/
Examples:
Use pretrained:
python ./run_dmrir_tadasae.py --test-only --checkpoint <path_to_weights>.pth
Train from scratch:
python ./run_dmrir_tadasae.py