Deep learning‐based semantic segmentation of interphase cells and debris from metaphase images

Abstract

Segmentation plays an essential role in the design of the automated karyotyping system (AKS). It is pivotal to segment interphase cells and other debris usually found in the input G metaphase images. The performance of AKSs is considerably less when interphase cells and debris are present in the input images. In this article, two semantic segmentation models are proposed. For this experiment, an annotated dataset is generated from the G banded metaphase images which are prepared at Regional Cancer Centre (RCC), Thiruvananthapuram, Kerala, India. Inspired by the results of UNet, a lighter version L‐UNet is developed and experimented with. It shows the validation IoU (Intersection over Union) of 0.9809 and F1‐score of 0.9903 on the RCC dataset and the test IoU of 0.9720 and F1‐score of 0.9858 on the CRCN‐NE dataset. As backbone semantic segmentation models are state of the art, an efficient model, Eff‐UNet, is also proposed here. In this model, EfficientNetB03 acts as the backbone that extracts powerful features and UNet acts as the decoder that predicts the segmentation map. It performs with the validation IoU of 0.9842 and F1‐score of 0.9920 on the RCC dataset and the test IoU of 0.7545 and F1‐score of 0.7778 on the CRCN‐NE dataset. To derive this model, 25 encoder–decoder architectures are evaluated with various top‐performing CNNs (convolutional neural networks) as encoders and segmentation networks as decoders. Results are further compared with various segmentation models and the best results are obtained from the proposed model.

A dual channel and spatial attention network for automatic spine segmentation of MRI images

“…It improved the performance of many medical image segmentations and detection tasks. Eff‐UNet 20 achieves higher DSC scores in interphase cells and debris segmentation tasks by improving the feature extraction module based on the U‐Net. MR‐UNet 21 adds a multi‐scale pyramid module to extract the characteristics of retinal vessels with different thicknesses.…”

Attention augmented residual autoencoder for efficient polyp segmentation

“…Chinnu et al 22 used Blind/Referenceless Image Spatial Quality Evaluator‐based approach to extract the slices having lung abnormalities from the dataset followed by a novel shallow CNN model to detect lung carcinoma from multimodality images. Remya et al 23 proposed L‐UNet and Eff‐UNet architectures for segmentation of G banded metaphase chromosome images. Zhang et al 24 applied transfer learning techniques utilizing features learned from big nonmedical datasets to detect and classify hyperplastic and adenomatous colorectal polyps.…”