File size: 11,320 Bytes
237cc3c
 
 
 
 
 
 
 
0858169
 
 
 
fd19e5e
0858169
 
 
 
 
fd19e5e
0858169
 
 
 
 
 
fd19e5e
 
0858169
fd19e5e
0858169
 
 
 
 
fd19e5e
0858169
 
 
 
 
 
 
fd19e5e
0858169
fd19e5e
0858169
fd19e5e
 
0858169
 
 
 
 
 
fd19e5e
0858169
 
fd19e5e
 
 
 
 
 
 
 
 
 
 
0858169
fd19e5e
 
 
 
 
 
 
 
 
0858169
fd19e5e
 
 
 
0858169
 
fd19e5e
0858169
fd19e5e
0858169
fd19e5e
 
 
 
 
 
 
 
 
0858169
 
fd19e5e
 
 
0858169
fd19e5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0858169
 
 
 
 
fd19e5e
 
 
 
 
 
 
 
0858169
fd19e5e
 
 
 
 
 
0858169
 
fd19e5e
0858169
fd19e5e
 
 
0858169
fd19e5e
 
0858169
fd19e5e
 
0858169
 
fd19e5e
0858169
 
 
 
 
fd19e5e
 
0858169
fd19e5e
 
0858169
fd19e5e
0858169
fd19e5e
 
 
0858169
fd19e5e
 
 
 
 
 
 
 
0858169
fd19e5e
0858169
 
fd19e5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0858169
 
fd19e5e
 
 
 
 
 
0858169
 
fd19e5e
 
 
0858169
fd19e5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c70a6c2
fd19e5e
 
0858169
 
fd19e5e
0858169
fd19e5e
 
0858169
fd19e5e
0858169
c70a6c2
 
 
 
 
0858169
c70a6c2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
# Patch missing audioop module (removed in Python 3.13) before importing gradio
import sys
import types

if 'audioop' not in sys.modules:
    _audioop = types.ModuleType('audioop')
    sys.modules['audioop'] = _audioop

import io
import json
import base64
import numpy as np
from PIL import Image, ImageDraw
import torch
import torch.nn as nn
from huggingface_hub import hf_hub_download
import gradio as gr

# ── HRNet-W32 ─────────────────────────────────────────────────────────────────

class BasicBlock(nn.Module):
    expansion = 1
    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super().__init__()
        self.conv1 = nn.Conv2d(inplanes, planes, 3, stride=stride, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(planes)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)
        self.downsample = downsample

    def forward(self, x):
        out = self.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))
        return self.relu(out + (self.downsample(x) if self.downsample else x))


class Bottleneck(nn.Module):
    expansion = 4
    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super().__init__()
        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
        self.bn1 = nn.BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, 3, stride=stride, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)
        self.conv3 = nn.Conv2d(planes, planes * 4, 1, bias=False)
        self.bn3 = nn.BatchNorm2d(planes * 4)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample

    def forward(self, x):
        out = self.relu(self.bn1(self.conv1(x)))
        out = self.relu(self.bn2(self.conv2(out)))
        out = self.bn3(self.conv3(out))
        return self.relu(out + (self.downsample(x) if self.downsample else x))


def make_layer(block, inplanes, planes, blocks, stride=1):
    downsample = None
    if stride != 1 or inplanes != planes * block.expansion:
        downsample = nn.Sequential(
            nn.Conv2d(inplanes, planes * block.expansion, 1, stride=stride, bias=False),
            nn.BatchNorm2d(planes * block.expansion)
        )
    layers = [block(inplanes, planes, stride, downsample)]
    for _ in range(1, blocks):
        layers.append(block(planes * block.expansion, planes))
    return nn.Sequential(*layers)


class FuseLayer(nn.Module):
    def __init__(self, num_branches, num_channels):
        super().__init__()
        self.num_branches = num_branches
        fuse = []
        for i in range(num_branches):
            row = []
            for j in range(num_branches):
                if j > i:
                    row.append(nn.Sequential(
                        nn.Conv2d(num_channels[j], num_channels[i], 1, bias=False),
                        nn.BatchNorm2d(num_channels[i]),
                        nn.Upsample(scale_factor=2 ** (j - i), mode='nearest')
                    ))
                elif j == i:
                    row.append(nn.Identity())
                else:
                    convs = []
                    for k in range(i - j):
                        inc = num_channels[j] if k == 0 else num_channels[i]
                        convs += [nn.Conv2d(inc, num_channels[i], 3, stride=2, padding=1, bias=False),
                                  nn.BatchNorm2d(num_channels[i])]
                        if k < i - j - 1:
                            convs.append(nn.ReLU(True))
                    row.append(nn.Sequential(*convs))
            fuse.append(nn.ModuleList(row))
        self.fuse = nn.ModuleList(fuse)
        self.relu = nn.ReLU(True)

    def forward(self, x):
        out = []
        for i in range(self.num_branches):
            y = x[0] if i == 0 else self.fuse[i][0](x[0])
            for j in range(1, self.num_branches):
                y = y + (x[j] if i == j else self.fuse[i][j](x[j]))
            out.append(self.relu(y))
        return out


class HRStage(nn.Module):
    def __init__(self, num_branches, block, num_blocks, num_channels):
        super().__init__()
        self.branches = nn.ModuleList([
            nn.Sequential(*[block(num_channels[i], num_channels[i]) for _ in range(num_blocks)])
            for i in range(num_branches)
        ])
        self.fuse = FuseLayer(num_branches, num_channels)

    def forward(self, x):
        x = [b(xi) for b, xi in zip(self.branches, x)]
        return self.fuse(x)


class HRNet(nn.Module):
    def __init__(self, num_joints=19):
        super().__init__()
        self.stem = nn.Sequential(
            nn.Conv2d(3, 64, 3, stride=2, padding=1, bias=False), nn.BatchNorm2d(64), nn.ReLU(True),
            nn.Conv2d(64, 64, 3, stride=2, padding=1, bias=False), nn.BatchNorm2d(64), nn.ReLU(True),
        )
        self.layer1 = make_layer(Bottleneck, 64, 64, 4)
        self.trans1 = nn.ModuleList([
            nn.Sequential(nn.Conv2d(256, 32, 3, padding=1, bias=False), nn.BatchNorm2d(32), nn.ReLU(True)),
            nn.Sequential(nn.Conv2d(256, 64, 3, stride=2, padding=1, bias=False), nn.BatchNorm2d(64), nn.ReLU(True)),
        ])
        self.stage2 = HRStage(2, BasicBlock, 4, [32, 64])
        self.trans2 = nn.Sequential(nn.Conv2d(64, 128, 3, stride=2, padding=1, bias=False), nn.BatchNorm2d(128), nn.ReLU(True))
        self.stage3 = nn.Sequential(*[HRStage(3, BasicBlock, 4, [32, 64, 128]) for _ in range(4)])
        self.trans3 = nn.Sequential(nn.Conv2d(128, 256, 3, stride=2, padding=1, bias=False), nn.BatchNorm2d(256), nn.ReLU(True))
        self.stage4 = nn.Sequential(*[HRStage(4, BasicBlock, 4, [32, 64, 128, 256]) for _ in range(3)])
        self.head = nn.Conv2d(32, num_joints, 1)

    def forward(self, x):
        x = self.stem(x)
        x = self.layer1(x)
        x = [t(x) for t in self.trans1]
        x = self.stage2(x)
        x = [x[0], x[1], self.trans2(x[1])]
        for m in self.stage3:
            x = m(x)
        x = [x[0], x[1], x[2], self.trans3(x[2])]
        for m in self.stage4:
            x = m(x)
        return self.head(x[0])


# ── Load weights ──────────────────────────────────────────────────────────────
print("Downloading model weights...")
model_path = hf_hub_download(
    repo_id="cwlachap/hrnet-cephalometric-landmark-detection",
    filename="best_model.pth"
)
checkpoint = torch.load(model_path, map_location="cpu", weights_only=False)
state_dict = checkpoint.get("model_state_dict", checkpoint.get("state_dict", checkpoint))
model = HRNet(num_joints=19)
missing, unexpected = model.load_state_dict(state_dict, strict=False)
print(f"Loaded. Missing: {len(missing)} | Unexpected: {len(unexpected)}")
model.eval()
print("Model ready.")

# ── Constants ─────────────────────────────────────────────────────────────────
LM_IDS = ['S', 'N', 'Or', 'Po', 'ANS', 'PNS', 'A', 'U1tip', 'L1tip', 'B',
           'Pog', 'Me', 'Gn', 'Go', 'Co', 'L1ap', 'U1ap', 'U6', 'L6']

LM_COLORS = {
    'S': '#58a6ff', 'N': '#58a6ff', 'Or': '#58a6ff', 'Po': '#58a6ff',
    'ANS': '#3fb950', 'PNS': '#3fb950', 'A': '#3fb950',
    'U1tip': '#3fb950', 'U1ap': '#3fb950', 'U6': '#3fb950',
    'B': '#f0883e', 'L1tip': '#f0883e', 'L1ap': '#f0883e',
    'Pog': '#f0883e', 'Me': '#f0883e', 'Gn': '#f0883e',
    'Go': '#f0883e', 'Co': '#f0883e', 'L6': '#f0883e'
}

INPUT_W, INPUT_H = 256, 320


# ── Helpers ───────────────────────────────────────────────────────────────────
def preprocess(pil_img):
    img = pil_img.convert('RGB').resize((INPUT_W, INPUT_H), Image.BILINEAR)
    arr = np.array(img, dtype=np.float32) / 255.0
    arr = (arr - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225]
    return torch.from_numpy(arr).permute(2, 0, 1).unsqueeze(0).float()


def heatmap_to_coords(hm_np, orig_w, orig_h):
    coords = {}
    nj, hh, hw = hm_np.shape
    for j in range(min(nj, len(LM_IDS))):
        hm = hm_np[j]
        idx = int(hm.argmax())
        py, px = divmod(idx, hw)
        if 1 <= px < hw - 1 and 1 <= py < hh - 1:
            px += 0.25 * np.sign(float(hm[py, px + 1] - hm[py, px - 1]))
            py += 0.25 * np.sign(float(hm[py + 1, px] - hm[py - 1, px]))
        x_norm = float(np.clip((px / hw) * (INPUT_W / orig_w), 0, 1))
        y_norm = float(np.clip((py / hh) * (INPUT_H / orig_h), 0, 1))
        coords[LM_IDS[j]] = {"x": round(x_norm, 4), "y": round(y_norm, 4), "confidence": 0.85}
    return coords


def run_detection(pil_img):
    orig_w, orig_h = pil_img.size
    tensor = preprocess(pil_img)
    with torch.no_grad():
        hm = model(tensor)[0].numpy()
    return heatmap_to_coords(hm, orig_w, orig_h)


# ── Gradio functions ──────────────────────────────────────────────────────────
def detect_visual(pil_img):
    if pil_img is None:
        return None, "{}"
    coords = run_detection(pil_img)
    out = pil_img.copy().convert("RGB")
    draw = ImageDraw.Draw(out)
    w, h = out.size
    r = max(4, w // 120)
    for lm_id, pt in coords.items():
        cx, cy = int(pt['x'] * w), int(pt['y'] * h)
        col = LM_COLORS.get(lm_id, '#ffffff')
        draw.ellipse([cx - r, cy - r, cx + r, cy + r], fill=col, outline='black', width=1)
        draw.text((cx + r + 2, cy - r), lm_id, fill=col)
    return out, json.dumps({"landmarks": coords}, indent=2)


def detect_api(image_b64: str) -> str:
    try:
        img_bytes = base64.b64decode(image_b64)
        pil_img = Image.open(io.BytesIO(img_bytes)).convert('RGB')
        coords = run_detection(pil_img)
        return json.dumps({"landmarks": coords})
    except Exception as e:
        return json.dumps({"error": str(e)})


# ── UI β€” single Blocks interface only, no nested Interface ────────────────────
with gr.Blocks(title="OrthoTimes Landmark Detection") as demo:
    gr.Markdown("## OrthoTimes QuickCephTool β€” HRNet Landmark Detection\nDetects 19 cephalometric landmarks automatically.")

    with gr.Row():
        img_in = gr.Image(type="pil", label="Upload lateral cephalogram")
        with gr.Column():
            img_out = gr.Image(type="pil", label="Detected landmarks")
            json_out = gr.Textbox(label="JSON output", lines=12)

    gr.Button("β–Ά Detect Landmarks").click(fn=detect_visual, inputs=img_in, outputs=[img_out, json_out])

    # API tab for programmatic access
    with gr.Tab("API"):
        b64_in  = gr.Textbox(label="Base64 encoded image", lines=3)
        b64_out = gr.Textbox(label="JSON result", lines=10)
        gr.Button("Detect (API)").click(fn=detect_api, inputs=b64_in, outputs=b64_out)

demo.launch(server_name="0.0.0.0", server_port=7860)