A Combined Deep CNN-LSTM Network for Chromosome Classification for Metaphase Selection

Chromosome karyotyping uses segmentation and classification of particular chromosome pictures generated from stained images obtained during the metaphase stage of cell cycle division. The segmenting and classification of single chromosomal pictures takes a considerable amount of manual labour hours in many clinics and labs. Deep learning models have been used to simplify this task recently, and the outcomes have been encouraging. Cytogeneticists use manual karyotyping to perform giemsa staining on chromosomes, which creates a sequence of lighter and darker bands. For chromosome categorization, this paper suggests CNN-LSTM which takes advantage of this band sequence characteristic. CNN-ResNet’s convolutional layers are input into LSTM, and the LSTM output is fed into an attention model, which labels the LSTM output sequences into one of 24 labels. LSTM is an end-to-end trainable model. Due to the attention mechanism that follows the recurrent layers, the network can discover a method to selectively pay attention to the band sequence and correlate it with distinct types of chromosomes. It is demonstrated that the suggested design surpasses baseline models constructed with classic deep convolutional neural networks and ResNet50 by about 3% Top-1 classification accuracy using a publicly accessible Bioimage chromosomal categorization database.