From 1041a66270a2f8ccf2746b6b3b92109c04b40c09 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E5=88=98=E8=88=AA=E5=AE=87?= <3364451258@qq.com>
Date: Fri, 8 May 2026 15:31:47 +0800
Subject: [PATCH] =?UTF-8?q?refactor:=20remove=5Fsmall=5Fobjects=E7=94=A8Ot?=
 =?UTF-8?q?su=E6=9B=BF=E4=BB=A3=E4=B8=AD=E4=BD=8D=E6=95=B025%?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

对连通域面积分布做Otsu自动找分界，不再拍脑袋定百分比
---
 .gitignore                  |  3 +++
 src/cDNA_gridding_simple.py | 32 +++++++++++++++++++++++++-------
 src/cDNA_segmentation.py    | 26 +++++++++++++++++++-------
 web/app.py                  | 13 ++++++++++---
 4 files changed, 57 insertions(+), 17 deletions(-)

diff --git a/.gitignore b/.gitignore
index 2a95f7e..f88bc1f 100644
--- a/.gitignore
+++ b/.gitignore
@@ -15,3 +15,6 @@ Thumbs.db
 
 # Obsidian
 .obsidian/
+
+# Flask
+.playwright-mcp/
\ No newline at end of file
diff --git a/src/cDNA_gridding_simple.py b/src/cDNA_gridding_simple.py
index bdf1cac..0e528a5 100644
--- a/src/cDNA_gridding_simple.py
+++ b/src/cDNA_gridding_simple.py
@@ -200,23 +200,41 @@ def remove_small_objects(binary: np.ndarray) -> np.ndarray:
     """
     自动去除小连通域（噪声）。
 
-    统计所有连通域面积的中位数，
-    小于中位数 25% 的视为噪声，直接剔除。
+    对连通域面积分布做 Otsu 阈值检测：
+    面积分布天然双峰——噪声区(几个像素) 和 真斑点区(几百像素)。
+    Otsu 自动找到两峰之间的最佳分界，小于该值的视为噪声。
     换图换分辨率都自动适应，不需要手动调参。
     """
     labeled, num = ndimage.label(binary)
     if num == 0:
-        return binary                  # 全黑，直接返回
+        return binary
 
     # 收集所有连通域的面积
-    areas = [int(np.sum(labeled == i)) for i in range(1, num + 1)]
-    median = np.median(areas)          # 面积中位数
-    min_size = max(1, int(median * 0.25))  # 中位数的25%，最少1像素
+    areas = np.array([int(np.sum(labeled == i)) for i in range(1, num + 1)])
+    if len(areas) < 2:
+        return binary
+
+    # 对面积数组做 Otsu（与像素 Otsu 完全相同的原理）
+    # 将面积值当作"灰度"，找到最小类内方差的分界点
+    best_T, best_cost, n_total = 0, float('inf'), len(areas)
+    for T in np.unique(areas):
+        small = areas[areas <= T]    # 候选噪声组
+        large = areas[areas > T]     # 候选真斑点组
+        w_s = len(small) / n_total
+        w_l = len(large) / n_total
+        if w_s == 0 or w_l == 0:
+            continue
+        cost = w_s * np.var(small) + w_l * np.var(large)
+        if cost < best_cost:
+            best_cost = cost
+            best_T = T
+
+    min_size = best_T  # Otsu 自动找到的面积分界线
 
     # 面积不达标的连通域整块置0
     result = binary.copy()
     for i in range(1, num + 1):
-        if areas[i - 1] < min_size:
+        if int(np.sum(labeled == i)) < min_size:
             result[labeled == i] = 0
     return result
 
diff --git a/src/cDNA_segmentation.py b/src/cDNA_segmentation.py
index 4f07e9e..c3b3860 100644
--- a/src/cDNA_segmentation.py
+++ b/src/cDNA_segmentation.py
@@ -207,21 +207,33 @@ def remove_small_objects(binary: np.ndarray) -> np.ndarray:
     """
     自动去除小连通域。
 
-    统计所有连通域的面积中位数，小于中位数 25% 的视为噪声，
-    自动剔除，不需要人设定阈值。
+    对连通域面积分布做 Otsu 阈值检测——面积天然双峰，
+    Otsu 自动找到噪声峰和真斑点峰之间的最佳分界，零人工参数。
     """
     labeled, num = ndimage.label(binary)
     if num == 0:
         return binary
 
-    # 统计每个连通域的面积
-    areas = [int(np.sum(labeled == i)) for i in range(1, num + 1)]
-    median_area = np.median(areas)
-    min_size = max(1, int(median_area * 0.25))  # 中位数的25%，最少1像素
+    areas = np.array([int(np.sum(labeled == i)) for i in range(1, num + 1)])
+    if len(areas) < 2:
+        return binary
+
+    best_T, best_cost, n_total = 0, float('inf'), len(areas)
+    for T_val in np.unique(areas):
+        small = areas[areas <= T_val]
+        large = areas[areas > T_val]
+        w_s = len(small) / n_total
+        w_l = len(large) / n_total
+        if w_s == 0 or w_l == 0:
+            continue
+        cost = w_s * np.var(small) + w_l * np.var(large)
+        if cost < best_cost:
+            best_cost = cost
+            best_T = T_val
 
     result = binary.copy()
     for i in range(1, num + 1):
-        if areas[i - 1] < min_size:
+        if int(np.sum(labeled == i)) < best_T:
             result[labeled == i] = 0
     return result
 
diff --git a/web/app.py b/web/app.py
index 282ec2b..bbeeb45 100644
--- a/web/app.py
+++ b/web/app.py
@@ -72,11 +72,18 @@ def keep_largest_object(binary):
 def remove_small_objects(binary):
     L, n = ndimage.label(binary)
     if n == 0: return binary
-    areas = [int(np.sum(L==i)) for i in range(1,n+1)]
-    minsz = max(1, int(np.median(areas)*0.25))
+    areas = np.array([int(np.sum(L==i)) for i in range(1,n+1)])
+    if len(areas) < 2: return binary
+    best_T, best_cost, n_total = 0, float('inf'), len(areas)
+    for T in np.unique(areas):
+        s, l = areas[areas<=T], areas[areas>T]
+        w_s, w_l = len(s)/n_total, len(l)/n_total
+        if w_s==0 or w_l==0: continue
+        cost = w_s*np.var(s) + w_l*np.var(l)
+        if cost < best_cost: best_cost, best_T = cost, T
     r = binary.copy()
     for i in range(1, n+1):
-        if areas[i-1] < minsz: r[L==i] = 0
+        if int(np.sum(L==i)) < best_T: r[L==i] = 0
     return r