CS-Net: Bridging Local and Temporal Features for Chromosomal Structural Abnormality Detection

Chromosomal structural abnormality detection is pivotal for diagnosing genetic disorders and advancing biomedical research. Existing methods primarily focus on instance-level features, making them suboptimal for identifying subtle structural abnormalities that manifest as localized banding pattern changes in chromosome images. Moreover, these methods overlook the temporal relationships inherent in chromosomes, which are critical for reasoning about structural abnormalities. These challenges are further compounded by the high appearance similarity among chromosomes, where traditional approaches such as deep learning-based classification or outlier analysis fail to effectively capture the nuanced variations. The lack of such modeling leaves existing methods unable to discover subtle and context-dependent banding changes from the instance-level representation of the chromosome. In this paper, we propose Chromosomal Sequence Net (CS-Net), a novel framework designed to address these challenges by explicitly modeling part-level features and spatial-temporal relationships in chromosome images. CS-Net introduces three key components: (1) C-Patcher, which segments the chromosome along its longitudinal axis into a sequence of image patches; (2) S-Encoder, which encodes the chromosome image sequence for part-level representation learning and models the spatial-temporal relationships among sequential patches using a newly designed predictive-positional loss; and (3) Ab-Detector, which leverages domain-specific medical priors to detect various types of structural abnormalities. Experiments on 76 distinct types of chromosomal abnormalities demonstrate that CS-Net achieves strong detection performance (e.g., a false positive rate of 0.116 and a false negative rate of 0.006). Compared with recent state-of-the-art methods, CS-Net remains competitive while covering a broader range of structural abnormalities.