A study of deep learning approaches for classification and detection chromosomes in metaphase images

Chromosome analysis is an important approach to detecting genetic diseases. However, the process of identifying chromosomes in metaphase images can be challenging and time-consuming. Therefore, it is important to use automatic methods for detecting chromosomes to aid diagnosis. This work proposes a study of deep learning approaches for classification and detection of chromosome in metaphase images. Furthermore, we propose a method for detecting chromosomes, which includes new stages for preprocessing and reducing false positives and false negatives. The proposed method is evaluated using 74 chromosome images in the metaphase stage, which were obtained from the CRCN-NE database, resulting in 2174 chromosome regions. We undertake three types of evaluation: segmentation; classification of cropped regions of chromosomes; and detection of chromosomes in the original images. For the segmentation analysis, we evaluated the Otsu, adaptive, fuzzy and fuzzy-adaptive methods. For classification and detection, we evaluated the following state-of-the-art algorithms: VGG16, VGG19, Inception v3, MobileNet, Xception, Sharma and MiniVGG. The classification results showed that the proposed approach, using segmented images, obtained better results than using RGB images. Furthermore, when analyzing deep learning approaches, the VGG16 algorithm obtained the best results, using fine tuning, with a sensitivity of 0.98, specificity of 0.99 and AUC of 0.955. The results also showed that the proposed negative reduction method increased sensitivity by 18%, while maintaining the specificity value. Deep learning methods have been proved to be efficient at detecting chromosomes, but preprocessing and post-processing are important to avoid false negatives. Therefore, using binary images and adding stages for reducing false positives and false negatives are necessary in order to increase the quality of the images of the chromosomes detected.