斑馬魚胚胎表皮細胞影像辨識與細胞分裂分析

Abstract

斑馬魚表皮細胞（superficial epidermal cells, SECs）為已分化且不再進入典型細胞週期的上皮細胞，但仍可透過非合成性分裂（asynthetic fission）進行增殖。結合 Palmskin／Brainbow 譜系追蹤系統所提供的穩定顏色遺傳特性，可藉由影像分析細胞群組大小與排列，以推測近期的細胞分裂活動。然而，從高解析度且具高噪訊與色彩變異的影像中自動化擷取此類資訊仍具挑戰。

本研究建立一套完整的斑馬魚 SEC 影像分析流程，包含：（1）單細胞分割模型、（2）細胞群組合併演算法，以及（3）灰階影像分割與機器學習／深度學習預測，首先以 736 張人工標註影像微調 Cellpose cyto3 模型，使平均 Dice score 由 0.7456 提升至 0.9171，經適度資料增強後進一步提升至 0.9369，此外，為探討低對比條件下的分割能力，提出感知亮度與 PCA 投影兩種灰階化方法，以僅依賴形態資訊進行模型訓練與評估。

在細胞群組重建方面，本研究整合鄰近性、CIE Lab 色彩距離以及 SEC 分裂次數的生物限制，設計細胞合併演算法，在 70 張測試影像上達到 0.078 的低錯誤率。

在細胞分裂辨識方面，比較 U-Net、DenseNet 與 Vision Transformer（ViT）等模型。結果顯示，在三分類任務中整體 F1 score 約為 0.49–0.53，而在二分類（是否發生分裂）任務中可達約 0.75，顯示細胞形態確實包含與分裂相關的可辨識特徵，整體而言DenseNet 在效能與穩定性之間取得最佳平衡。

本研究結合生物先驗知識與深度學習技術，建立了一套可擴展的 SEC 自動化影像分析流程，能有效重建細胞族群結構並量化分裂行為，為未來整合時間序列、三維影像與分子標記之研究奠定基礎。

Zebrafish superficial epidermal cells (SECs) are differentiated epithelial cells that no longer enter the canonical cell cycle. However, they can still proliferate through asynthetic fission. When combined with the stable color inheritance provided by the Palmskin/Brainbow lineage-tracing system, image analysis of cell group size and spatial arrangement can be used to infer recent cell division events. Nevertheless, automatically extracting such information from high-resolution images with substantial noise and color variation remains challenging.

In this study, we developed a complete SEC image analysis pipeline. The pipeline includes (1) a single-cell segmentation model, (2) a cell-group merging algorithm, and (3) grayscale-based segmentation with machine learning and deep learning prediction. First, the Cellpose cyto3 model was fine-tuned using 736 manually annotated images. This step improved the average Dice score from 0.7456 to 0.9171. After applying appropriate data augmentation, the Dice score further increased to 0.9369. To evaluate segmentation performance under low-contrast conditions, two grayscale conversion methods were introduced: perceived luminance transformation and PCA projection. These methods allow the model to be trained and evaluated using morphological information only.

For cell-group reconstruction, this study integrated spatial proximity, CIE Lab color distance, and biological constraints on SEC division events to design a cell-merging algorithm. The algorithm achieved a low error rate of 0.078 on 70 test images.

For cell division identification, several models were compared, including U-Net, DenseNet, and Vision Transformer (ViT). In the three-class classification task, the overall F1 score ranged from 0.49 to 0.53. In the binary classification task (division vs. non-division), the F1 score reached approximately 0.75. These results indicate that cell morphology contains identifiable features related to division behavior. Overall, DenseNet achieved the best balance between performance and stability.

This study integrates biological prior knowledge with deep learning techniques to establish a scalable SEC automated image analysis pipeline. The framework can effectively reconstruct cell population structures and quantify division behavior. It also provides a foundation for future studies integrating time-series data, three-dimensional imaging, and molecular markers.