Enhanced label constrained contrastive learning for chromosome optical microscopic image classification

Abstract

Karyotyping analysis, an indispensable means of diagnosing genetic diseases in clinical practice, entails a highly intricate process of chromosome classification. Automatic chromosome classification, in particular, poses a daunting challenge, impeding the acceleration of the diagnosis speed of karyotyping analysis. Hence, we propose a newfangled method of enhanced label-constrained contrastive learning to sensitively discriminate chromosomes. Our method incorporates a multi-scale feature perception aggregation encoder and a linear classifier and operates in two stages. The first stage involves the utilization of an encoder and projection head for chromosome representation learning. We optimally exploit label information and conduct contrastive learning on positive samples of the same category of chromosomes. Further, we fortify local irregularities by means of CutMix image augmentation and extract global and local features via multi-scale feature modules. The input perception aggregation module establishes feature dependencies and introduces a supervised loss to minimize intra-class distance and enlarge inter-class distance in the representation space. Finally, we map the learned representation to the Multi-Layer Perception (MLP) projection head for chromosome representation learning. In the second stage, we freeze the encoder, discard the projection head, and fine-tune the chromosome classification network by adding a classification layer that assigns each chromosome to a type in each patient case. We evaluate our method thoroughly on two public datasets and a private dataset, and the detailed results indicate that our proposed enhanced label-constrained contrastive learning attains the highest average accuracy of 99.272%. Our method outperforms state-of-the-art approaches, manifesting the effectiveness of our enhanced label-supervised contrastive learning, multi-scale feature integration, and feature dependency relationships. Additionally, it is also amenable to other medical image-related tasks.