Step 5: Image Quantification

概覽

一張美的影像 ≠ 可量化的影像

顯微鏡影像進入 ImageJ/Fiji (Schindelin 2012, Nat Methods) 後，pixel 值受 (1) 偵測器增益、(2) 雷射功率、(3) 樣品染色批次、(4) 顯示窗口 LUT 影響。任一項在實驗組與對照組間不同就會偽造差異。QUAREP-LiMi 自 2020 年起推動光顯微鏡品保標準，4DN-BINA-OME 則制定影像元數據規範。

2025 年 Cellpose3 (Stringer & Pachitariu, Nat Methods) 整合一鍵影像修復，搭配 StarDist (Schmidt 2018)、QuPath (Bankhead 2017) 與 CellProfiler 4 (Stirling 2021) 成為主流四大工具。

Microscopy images entering Fiji/ImageJ (Schindelin 2012, Nat Methods) carry pixel values affected by (1) detector gain, (2) laser power, (3) staining batch, (4) display LUT. Any of these differing between groups invents fake differences. QUAREP-LiMi drives microscopy QA standards since 2020; 4DN-BINA-OME defines image metadata.

By 2025, Cellpose3 (Stringer & Pachitariu, Nat Methods) adds one-click image restoration; together with StarDist (Schmidt 2018), QuPath (Bankhead 2017), and CellProfiler 4 (Stirling 2021), these form the dominant toolkit.

🚨

絕對禁忌：(1) 用 8-bit JPEG 量化（壓縮失真）；(2) 在 Brightness/Contrast 調整後的影像上量化；(3) 不同樣品用不同曝光時間。原始 16-bit TIFF + 一致採集參數是定量的最低門檻。Absolute don'ts: (1) Quantifying on 8-bit JPEG (lossy compression); (2) Quantifying after Brightness/Contrast adjustment; (3) Different exposure between samples. Raw 16-bit TIFF + identical acquisition settings is the floor.

核心概念

一、五個必懂概念

📐 ROI

手動勾選或自動分割（Cellpose/StarDist）。報告時需描述：是否盲法、最低細胞數、視野數。

Manual or automated (Cellpose/StarDist). Always report: blinded? min cell count? number of fields?

🌑 背景扣除

Process > Subtract Background（rolling ball，半徑略大於最大物件）。共定位前必扣，否則背景拉升 Pearson r。

Process > Subtract Background (rolling ball, radius slightly > largest object). Required before colocalization or background inflates Pearson r.

🎨 Pearson r

強度線性相關 (−1~+1)；對背景敏感、忽略「比例」資訊。適合分布相似度判斷。

Linear intensity correlation (−1 to +1); background-sensitive; ignores "overlap fraction". Use for distribution similarity.

🎯 Manders M1/M2

「A 訊號中與 B 重疊比例」。需先設閾值，建議用 Costes automatic thresholding 避免主觀。

"Fraction of A overlapping with B." Requires threshold; use Costes automatic thresholding to avoid bias.

互動模擬

背景對共定位的影響

調整背景強度與閾值，觀察 Pearson r 與 Manders M1 變化。關鍵：不扣背景時 Pearson r 會被假性拉高（因為背景在兩通道都「同步」）。

Adjust background and threshold — see Pearson r and Manders M1 shift. Key insight: without background subtraction, Pearson r is falsely inflated (background co-varies across channels).

背景強度20

閾值50

——

X：通道 A | Y：通道 B

程式碼

二、影像量化程式碼

// Fiji macro：批次量化資料夾內所有 TIFF
dir = getDirectory("input");
files = getFileList(dir);
setOption("ExpandableArrays", true);
for (i=0; i<files.length; i++) {
  open(dir+files[i]);
  run("Subtract Background...", "rolling=50");
  setAutoThreshold("Otsu dark");
  run("Analyze Particles...", "size=20-Infinity display");
  close();
}
// R: 後處理 ImageJ 輸出的 Results.csv
library(tidyverse)
res <- read_csv("Results.csv") %>%
  mutate(image = str_extract(Label, "[^:]+")) %>%
  group_by(image) %>%
  summarise(n=n(), meanInt=mean(Mean), totalArea=sum(Area))

import numpy as np, tifffile, pandas as pd
from skimage import filters, morphology, measure
from scipy.stats import pearsonr

img = tifffile.imread("sample.tif")        # shape (C, Y, X)
chA, chB = img[0], img[1]

# Rolling ball background subtraction
from skimage.restoration import rolling_ball
chA_bg = chA - rolling_ball(chA, radius=50)
chB_bg = chB - rolling_ball(chB, radius=50)

# 閾值（Otsu）後計算 Pearson 與 Manders
tA = chA_bg > filters.threshold_otsu(chA_bg)
tB = chB_bg > filters.threshold_otsu(chB_bg)
mask = tA | tB
pearson_r, _ = pearsonr(chA_bg[mask].ravel(), chB_bg[mask].ravel())
M1 = (chA_bg * tB).sum() / chA_bg.sum()
M2 = (chB_bg * tA).sum() / chB_bg.sum()
print(f"Pearson r = {pearson_r:.3f}, M1 = {M1:.3f}, M2 = {M2:.3f}")

# Cellpose3 一鍵分割（2025）
from cellpose import models, denoise
model = models.CellposeModel(gpu=True, model_type="cyto3")
masks, flows, _ = model.eval(chA_bg, diameter=30)
props = measure.regionprops_table(masks, intensity_image=chA_bg, properties=("area","mean_intensity"))
pd.DataFrame(props).to_csv("cells.csv", index=False)

常見陷阱

三、reviewer 必問清單

❌ 代表性圖片 = 分析

一張代表圖只是定性。量化必須 ≥ 3 個獨立樣本、每樣本 ≥ 3 個視野、每視野 ≥ 20 個物件。

One representative image is qualitative. Quantify ≥3 independent samples × ≥3 fields × ≥20 objects.

❌ 不同曝光量化

所有比較組必須用相同採集設定。亮度/對比可在展示時不同，但量化階段不可。

All compared groups must use identical acquisition. Display brightness/contrast can differ for figures, but never for quantification.

❌ ROI 選擇偏差

「選好看的細胞」是嚴重 confounding。盲法或自動分割是底線。

"Picking nice cells" is severe confounding. Blinding or automated segmentation is mandatory.

❌ 無 scale bar / metadata

4DN-BINA-OME 要求 OME-TIFF 帶物理像素大小、通道名、雷射波長。Fiji 的「Bio-Formats」可保存。

4DN-BINA-OME requires OME-TIFF with physical pixel size, channel name, laser wavelength. Save via Fiji's "Bio-Formats."

🎯 章末小測驗

1. 量化共定位前必須做？

轉 JPEG

背景扣除

3D 重建

增亮度

2. 量化用的最低影像規格？

8-bit JPEG

16-bit TIFF

PNG with transparency

GIF animation

3. Pearson r 與 Manders M1 差別？

只是符號

線性相關 vs 重疊比例

相同

Manders 限 RGB