Chromosome Straightening via Disentangled Representations: Exploring Semantic Trajectories in GAN’s Latent Space

In the field of chromosome karyotype analysis, performing straightening preprocessing on chromosomes is a key process to increase the accuracy of chromosome identification. Previous studies have typically relied on geometric algorithms; however, during the straightening process, external perturbations caused by geometric factors often result in the distortion or deformation of chromosome banding patterns, leading to the loss of important feature information and the lack of clarity in details. This, in turn, reduces the model’s representation capability and generalization performance. In this paper, we propose a novel disentangled representation learning framework aimed at exploring the semantic disentanglement trajectories within generative models to generate chromosome images with purer straightening semantics. Our framework consists of four main components: the Generator(G), the Semantic Trajectory Explorer(STE), the Disentangled Encoder(DE), and the Classifier(C). The G is responsible for generating chromosome images, the STE explores disentangled semantic trajectories within the latent space of the generative model, the DE maps chromosome images into the disentangled feature space, and the C predicts the chromosome type based on these features. The framework employs a disentanglement loss to force the STE to find disentangled semantic trajectories within the latent space of the G and uses classification loss to ensure that content information remains unchanged along the semantic trajectory, thereby achieving chromosome straightening without altering karyotype features. Based on the experimental results, our method performs well in terms of the Aspect Ratio Change of the bounding box(Bounding Box ARC), Frechet Inception Distance (FID), and Downstream Classification Accuracy (DCA).