Leveraging Latent Space and Radiomics for Multi-Label Classification of Chest CT Scans

Axel Bessy, Thomas Barba, Alexandre Meyer, Hamid Ladjal, Antoine Richard, Mathieu Lefort, Simon Achard

PFIA 2025, June 2025

Abstract

In this article, we propose two approaches for multi-label diagnosis from chest CT images. The first, called Seg2Clf, leverages the latent space of the encoder from a pre-trained segmentation model to extract discriminative representations for classification. The second approach relies on the extraction and integration of clinical radiomic features, allowing prediction refinement using a k-NN classifier. These two complementary methods are part of a broader strategy aimed at developing automated systems capable of simultaneously identifying multiple thoracic pathologies with improved accuracy.

Bibtex

@inproceedings{bessy2025,
  TITLE = {{Leveraging Latent Space and Radiomics for Multi-Label Classification of Chest CT Scans}},
  AUTHOR = {Bessy, Axel and Barba, Thomas and Meyer, Alexandre and Ladjal, Hamid and Richard, Antoine and Lefort, Mathieu and Achard, Simon},
  BOOKTITLE = {PFIA 2025},
  ADDRESS = {Dijon, France},
  YEAR = {2025},
  MONTH = {June},
  KEYWORDS = {Multi-label classification ; Computed Tomography ; Transfer learning ; Radiomics ; Medical imaging},
  abstract = {In this article, we propose two approaches for multi-label
diagnosis from chest CT images. The first, called Seg2Clf,
leverages the latent space of the encoder from a pre-trained
segmentation model to extract discriminative representations
for classification. The second approach relies on the
extraction and integration of clinical radiomic features, allowing
prediction refinement using a k-NN classifier. These
two complementary methods are part of a broader strategy
aimed at developing automated systems capable of simultaneously
identifying multiple thoracic pathologies with
improved accuracy.}
}