add app

app.py

@@ -0,0 +1,94 @@
+import os
+import numpy as np
+import gradio as gr
+import cv2
+from cellpose import models
+from matplotlib.colors import hsv_to_rgb
+import os
+
+try:
+    model = models.CellposeModel(gpu=False, pretrained_model="cyto3")
+except Exception as e:
+    print(f"Error loading model: {e}")
+    exit(1)
+
+def plot_flows(y):
+    Y = (np.clip(normalize99(y[0][0]),0,1) - 0.5) * 2
+    X = (np.clip(normalize99(y[1][0]),0,1) - 0.5) * 2
+    H = (np.arctan2(Y, X) + np.pi) / (2*np.pi)
+    S = normalize99(y[0][0]**2 + y[1][0]**2)
+    HSV = np.concatenate((H[:,:,np.newaxis], S[:,:,np.newaxis], S[:,:,np.newaxis]), axis=-1)
+    HSV = np.clip(HSV, 0.0, 1.0)
+    flow = (hsv_to_rgb(HSV) * 255).astype(np.uint8)
+    return flow
+
+def plot_outlines(masks):
+    outpix = []
+    contours, hierarchy = cv2.findContours(masks.astype(np.int32), mode=cv2.RETR_FLOODFILL, method=cv2.CHAIN_APPROX_SIMPLE)
+    for c in range(len(contours)):
+        pix = contours[c].astype(int).squeeze()
+        if len(pix)>4:
+            peri = cv2.arcLength(contours[c], True)
+            approx = cv2.approxPolyDP(contours[c], 0.001, True)[:,0,:]
+            outpix.append(approx)
+    return outpix
+
+def plot_overlay(img, masks):
+    img = normalize99(img.astype(np.float32).mean(axis=-1))
+    img -= img.min()
+    img /= img.max()
+    HSV = np.zeros((img.shape[0], img.shape[1], 3), np.float32)
+    HSV[:,:,2] = np.clip(img*1.5, 0, 1.0)
+    for n in range(int(masks.max())):
+        ipix = (masks==n+1).nonzero()
+        HSV[ipix[0],ipix[1],0] = np.random.rand()
+        HSV[ipix[0],ipix[1],1] = 1.0
+    RGB = (hsv_to_rgb(HSV) * 255).astype(np.uint8)
+    return RGB
+
+def normalize99(img):
+    X = img.copy()
+    X = (X - np.percentile(X, 1)) / (np.percentile(X, 99) - np.percentile(X, 1))
+    return X
+
+def image_resize(img, resize=224):
+    ny,nx = img.shape[:2]
+    if np.array(img.shape).max() > resize:
+        if ny>nx:
+            nx = int(nx/ny * resize)
+            ny = resize
+        else:
+            ny = int(ny/nx * resize)
+            nx = resize
+        shape = (nx,ny)
+        img = cv2.resize(img, shape)
+    img = img.astype(np.uint8)
+    return img
+
+def cellpose_segment(img):
+    img_input = image_resize(img)
+    masks, flows, _ = model.eval(img_input)
+    flows = flows[0]
+    # masks = np.zeros(img.shape[:2])
+    # flows = np.zeros_like(img)
+    target_size = (img_input.shape[1], img_input.shape[0])
+    if (target_size[0]!=img.shape[1] and target_size[1]!=img.shape[0]):
+        # scale it back to keep the orignal size
+        masks = cv2.resize(masks.astype('uint16'), target_size, interpolation=cv2.INTER_NEAREST).astype('uint16')
+        flows = cv2.resize(flows.astype('float32'), target_size).astype('uint8')
+
+    outpix = plot_outlines(masks)
+    overlay = plot_overlay(img, masks)
+
+    return outpix, overlay, flows, masks 
+
+# Gradio Interface
+iface = gr.Interface(
+    fn=cellpose_segment, 
+    inputs="image", 
+    outputs=["image", "image", "image", "image"],
+    title="cellpose segmentation",
+    description="upload an image, then cellpose will segment it at a max size of 224x224"
+)
+
+iface.launch()