ecSegCls: Deep learning-based method for detecting extrachromosomal DNAs in both interphase and metaphase cancer cells

Extrachromosomal DNA (ecDNA) is an acentric circular DNA element that derives from but exists independently of chromosomes. EcDNAs often carry oncogenes with high copy numbers, contributing to tumor heterogeneity, and poor patient outcomes. Whole genome sequencing provides genome-wide detection of ecDNAs but lacks spatial resolution, while imaging methods such as fluorescence in situ hybridization (FISH) capture spatial context but rely on labor-intensive manual annotation by experts. Existing tools for automated detection of ecDNAs from FISH, such as ecSeg, partially address this limitation but remain restricted to metaphase cells, demonstrating modest classification performance. Here, we present ecSegCls, an automated pipeline for segmenting and classifying ecDNA in FISH and DAPI images with high accuracy. Our proposed method exploits both deep learning-based segmentation model and its extracted features, as well as XGBoost-based classification model, leading to the pipeline of segmenting nuclei, chromosomes, and ecDNA regions, and predicting the presence of ecDNA in both metaphase and interphase cells. Data augmentation and noise simulation were used for improved robustness and segmentation-derived features were used for training an XGBoost classifier. Ablation studies further identified key predictive features, enhancing interpretability. Using a public dataset of 483 FISH images from cancer cell lines as a model training set and a dataset of 776 FISH images internally generated 8 cancer cell lines as a classification set, we assessed and compared the performance of our model with those of previously published architectures, including UNet, UNet++, DeepLabV3+, Swin UNet, FATNet, HiFormer, DAEFormer, and ecSeg. Our proposed ecSegCls has achieved remarkable qualitative and quantitative performance, yielding high performance on diverse regions with diverse metrics in both segmentation and classification benchmarks, thus establishing ecSegCls as a robust and scalable automated framework for accurate ecDNA detection, advancing imaging-based research and clinical applications with ecDNA.