Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,582 +1,256 @@
|
|
1 |
#!/usr/bin/env python
|
2 |
-
# -*- coding: utf-8 -*-
|
3 |
-
|
4 |
"""
|
5 |
-
|
|
|
6 |
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
2
|
11 |
-
3. Made SAM-2 functionality optional with graceful fallback
|
12 |
-
4. Added installation instructions and requirements check
|
13 |
"""
|
|
|
14 |
|
15 |
-
|
16 |
-
|
17 |
-
# ---------------------------------------------------------------------
|
18 |
-
import os
|
19 |
import sys
|
20 |
-
import
|
21 |
-
import tempfile
|
22 |
import subprocess
|
23 |
-
import
|
24 |
-
from
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
warnings.filterwarnings("ignore", message=r".*upsample_bicubic2d.*")
|
29 |
|
30 |
-
|
31 |
-
|
32 |
-
# ---------------------------------------------------------------------
|
33 |
-
import torch
|
34 |
import numpy as np
|
35 |
-
from
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
try:
|
44 |
-
print("[SAM-2 Debug] Attempting to import SAM-2 modules...")
|
45 |
from sam2.build_sam import build_sam2
|
46 |
-
from sam2.automatic_mask_generator import SAM2AutomaticMaskGenerator
|
47 |
-
print("[SAM-2 Debug] Successfully imported SAM-2 modules")
|
48 |
return True, "SAM-2 already available"
|
49 |
-
except ImportError
|
50 |
-
|
51 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
try:
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
print("[SAM-2 Debug] Repository cloned successfully")
|
61 |
-
|
62 |
-
# Install SAM-2
|
63 |
-
print("[SAM-2 Debug] Installing SAM-2...")
|
64 |
-
original_dir = os.getcwd()
|
65 |
-
os.chdir("segment-anything-2")
|
66 |
-
subprocess.run([sys.executable, "-m", "pip", "install", "-e", "."], check=True)
|
67 |
-
os.chdir(original_dir)
|
68 |
-
print("[SAM-2 Debug] Installation completed")
|
69 |
-
|
70 |
-
# Add to Python path
|
71 |
-
sam2_path = os.path.abspath("segment-anything-2")
|
72 |
-
if sam2_path not in sys.path:
|
73 |
-
sys.path.insert(0, sam2_path)
|
74 |
-
print(f"[SAM-2 Debug] Added {sam2_path} to Python path")
|
75 |
-
|
76 |
-
# Try importing again
|
77 |
-
print("[SAM-2 Debug] Attempting to import SAM-2 modules again...")
|
78 |
-
from sam2.build_sam import build_sam2
|
79 |
-
from sam2.automatic_mask_generator import SAM2AutomaticMaskGenerator
|
80 |
-
print("[SAM-2 Debug] Successfully imported SAM-2 modules after installation")
|
81 |
-
return True, "SAM-2 installed successfully"
|
82 |
-
|
83 |
-
except Exception as e:
|
84 |
-
print(f"[SAM-2 Debug] Installation failed: {str(e)}")
|
85 |
-
print(f"[SAM-2 Debug] Error type: {type(e).__name__}")
|
86 |
-
return False, f"SAM-2 installation failed: {e}"
|
87 |
|
88 |
-
# Check SAM-2 availability
|
89 |
SAM2_AVAILABLE, SAM2_STATUS = check_and_install_sam2()
|
90 |
print(f"SAM-2 Status: {SAM2_STATUS}")
|
91 |
-
|
92 |
-
# =============================================================================
|
93 |
-
# SAM-2 imports (conditional)
|
94 |
-
# =============================================================================
|
95 |
if SAM2_AVAILABLE:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
try:
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
# =============================================================================
|
105 |
-
# Qwen-VLM imports & helper
|
106 |
-
# =============================================================================
|
107 |
-
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
|
108 |
-
from qwen_vl_utils import process_vision_info
|
109 |
-
|
110 |
-
# =============================================================================
|
111 |
-
# CheXagent imports
|
112 |
-
# =============================================================================
|
113 |
-
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
|
114 |
-
|
115 |
-
# ---------------------------------------------------------------------
|
116 |
-
# Devices
|
117 |
-
# ---------------------------------------------------------------------
|
118 |
-
def get_device():
|
119 |
-
if torch.cuda.is_available():
|
120 |
-
return torch.device("cuda")
|
121 |
-
if torch.backends.mps.is_available():
|
122 |
-
return torch.device("mps")
|
123 |
-
return torch.device("cpu")
|
124 |
-
|
125 |
-
# =============================================================================
|
126 |
-
# Qwen-VLM model & agent
|
127 |
-
# =============================================================================
|
128 |
-
_qwen_model = None
|
129 |
-
_qwen_processor = None
|
130 |
-
_qwen_device = None
|
131 |
-
|
132 |
-
def load_qwen_model_and_processor(hf_token=None):
|
133 |
-
global _qwen_model, _qwen_processor, _qwen_device
|
134 |
-
if _qwen_model is None:
|
135 |
-
_qwen_device = "mps" if torch.backends.mps.is_available() else "cpu"
|
136 |
-
print(f"[Qwen] loading model on {_qwen_device}")
|
137 |
-
auth_kwargs = {"use_auth_token": hf_token} if hf_token else {}
|
138 |
-
_qwen_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
139 |
-
"Qwen/Qwen2.5-VL-3B-Instruct",
|
140 |
-
trust_remote_code=True,
|
141 |
-
attn_implementation="eager",
|
142 |
-
torch_dtype=torch.float32,
|
143 |
-
low_cpu_mem_usage=True,
|
144 |
-
device_map=None,
|
145 |
-
**auth_kwargs,
|
146 |
-
).to(_qwen_device)
|
147 |
-
_qwen_processor = AutoProcessor.from_pretrained(
|
148 |
-
"Qwen/Qwen2.5-VL-3B-Instruct",
|
149 |
-
trust_remote_code=True,
|
150 |
-
**auth_kwargs,
|
151 |
-
)
|
152 |
-
return _qwen_model, _qwen_processor, _qwen_device
|
153 |
-
|
154 |
-
class MedicalVLMAgent:
|
155 |
-
"""Light wrapper around Qwen-VLM with an optional image."""
|
156 |
-
|
157 |
-
def __init__(self, model, processor, device):
|
158 |
-
self.model = model
|
159 |
-
self.processor = processor
|
160 |
-
self.device = device
|
161 |
-
self.system_prompt = (
|
162 |
-
"You are a medical information assistant with vision capabilities.\n"
|
163 |
-
"Disclaimer: I am not a licensed medical professional. "
|
164 |
-
"The information provided is for reference only and should not be taken as medical advice."
|
165 |
-
)
|
166 |
-
|
167 |
-
def run(self, user_text: str, image: Image.Image | None = None) -> str:
|
168 |
-
messages = [
|
169 |
-
{"role": "system", "content": [{"type": "text", "text": self.system_prompt}]}
|
170 |
-
]
|
171 |
-
user_content = []
|
172 |
-
if image is not None:
|
173 |
-
tmp = f"/tmp/{uuid.uuid4()}.png"
|
174 |
-
image.save(tmp)
|
175 |
-
user_content.append({"type": "image", "image": tmp})
|
176 |
-
user_content.append({"type": "text", "text": user_text or "Please describe the image."})
|
177 |
-
messages.append({"role": "user", "content": user_content})
|
178 |
-
|
179 |
-
prompt_text = self.processor.apply_chat_template(
|
180 |
-
messages, tokenize=False, add_generation_prompt=True
|
181 |
-
)
|
182 |
-
img_inputs, vid_inputs = process_vision_info(messages)
|
183 |
-
inputs = self.processor(
|
184 |
-
text=[prompt_text],
|
185 |
-
images=img_inputs,
|
186 |
-
videos=vid_inputs,
|
187 |
-
padding=True,
|
188 |
-
return_tensors="pt",
|
189 |
-
).to(self.device)
|
190 |
-
|
191 |
-
with torch.no_grad():
|
192 |
-
out = self.model.generate(**inputs, max_new_tokens=128)
|
193 |
-
trimmed = out[0][inputs.input_ids.shape[1] :]
|
194 |
-
return self.processor.decode(trimmed, skip_special_tokens=True).strip()
|
195 |
-
|
196 |
-
# =============================================================================
|
197 |
-
# SAM-2 model + AutomaticMaskGenerator (final minimal version)
|
198 |
-
# =============================================================================
|
199 |
-
import os
|
200 |
-
import numpy as np
|
201 |
-
from PIL import Image, ImageDraw
|
202 |
-
from sam2.build_sam import build_sam2
|
203 |
-
from sam2.automatic_mask_generator import SAM2AutomaticMaskGenerator
|
204 |
-
|
205 |
-
def initialize_sam2():
|
206 |
-
# These two files are already in your repo
|
207 |
-
CKPT = "checkpoints/sam2.1_hiera_large.pt" # ≈2.7 GB
|
208 |
-
CFG = "configs/sam2.1/sam2.1_hiera_l.yaml"
|
209 |
-
|
210 |
-
# One chdir so Hydra's search path starts inside sam2/sam2/
|
211 |
-
os.chdir("sam2/sam2")
|
212 |
-
|
213 |
-
device = get_device()
|
214 |
-
print(f"[SAM-2] building model on {device}")
|
215 |
-
|
216 |
-
sam2_model = build_sam2(
|
217 |
-
CFG, # relative to sam2/sam2/
|
218 |
-
CKPT, # relative after chdir
|
219 |
-
device=device,
|
220 |
-
apply_postprocessing=False,
|
221 |
-
)
|
222 |
-
|
223 |
-
mask_gen = SAM2AutomaticMaskGenerator(
|
224 |
-
model=sam2_model,
|
225 |
-
points_per_side=32,
|
226 |
-
pred_iou_thresh=0.86,
|
227 |
-
stability_score_thresh=0.92,
|
228 |
-
crop_n_layers=0,
|
229 |
-
)
|
230 |
-
return sam2_model, mask_gen
|
231 |
-
|
232 |
-
# ---------------------- build once ----------------------
|
233 |
-
try:
|
234 |
-
_sam2_model, _mask_generator = initialize_sam2()
|
235 |
-
print("[SAM-2] Successfully initialized!")
|
236 |
-
except Exception as e:
|
237 |
-
print(f"[SAM-2] Failed to initialize: {e}")
|
238 |
-
_sam2_model, _mask_generator = None, None
|
239 |
-
|
240 |
-
def automatic_mask_overlay(image_np: np.ndarray) -> np.ndarray:
|
241 |
-
"""Generate masks and alpha-blend them on top of the original image."""
|
242 |
-
if _mask_generator is None:
|
243 |
-
raise RuntimeError("SAM-2 mask generator not initialized")
|
244 |
-
|
245 |
-
anns = _mask_generator.generate(image_np)
|
246 |
-
if not anns:
|
247 |
-
return image_np
|
248 |
-
|
249 |
-
overlay = image_np.copy()
|
250 |
-
if overlay.ndim == 2: # grayscale → RGB
|
251 |
-
overlay = np.stack([overlay] * 3, axis=2)
|
252 |
-
|
253 |
-
for ann in sorted(anns, key=lambda x: x["area"], reverse=True):
|
254 |
-
m = ann["segmentation"]
|
255 |
-
color = np.random.randint(0, 255, 3, dtype=np.uint8)
|
256 |
-
overlay[m] = (overlay[m] * 0.5 + color * 0.5).astype(np.uint8)
|
257 |
-
|
258 |
-
return overlay
|
259 |
-
|
260 |
-
def tumor_segmentation_interface(image: Image.Image | None):
|
261 |
-
if image is None:
|
262 |
-
return None, "Please upload an image."
|
263 |
-
|
264 |
-
if _mask_generator is None:
|
265 |
-
return None, "SAM-2 not properly initialized. Check the console for errors."
|
266 |
-
|
267 |
-
try:
|
268 |
-
img_np = np.array(image.convert("RGB"))
|
269 |
-
out_np = automatic_mask_overlay(img_np)
|
270 |
-
n_masks = len(_mask_generator.generate(img_np))
|
271 |
-
return Image.fromarray(out_np), f"{n_masks} masks found."
|
272 |
except Exception as e:
|
273 |
-
|
274 |
-
|
275 |
-
#
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
return None, "Please upload an image."
|
282 |
-
|
283 |
try:
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
305 |
-
|
306 |
-
|
307 |
-
|
308 |
-
|
309 |
-
|
310 |
-
|
311 |
-
|
312 |
-
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
|
320 |
-
|
321 |
-
|
322 |
-
|
323 |
-
|
324 |
-
|
325 |
-
|
326 |
-
|
327 |
-
|
328 |
-
|
329 |
-
|
330 |
-
|
331 |
-
|
332 |
-
|
333 |
-
|
334 |
-
|
335 |
-
|
336 |
-
if torch.cuda.is_available():
|
337 |
-
print("[CheXagent] Converting to half precision for GPU")
|
338 |
-
chex_model = chex_model.half()
|
339 |
else:
|
340 |
-
|
341 |
-
chex_model = chex_model.float()
|
342 |
-
|
343 |
-
chex_model.eval()
|
344 |
-
CHEXAGENT_AVAILABLE = True
|
345 |
-
print("[CheXagent] Initialization complete")
|
346 |
-
except Exception as e:
|
347 |
-
print(f"[CheXagent] Initialization failed: {str(e)}")
|
348 |
-
print(f"[CheXagent] Error type: {type(e).__name__}")
|
349 |
-
CHEXAGENT_AVAILABLE = False
|
350 |
-
chex_tok, chex_model = None, None
|
351 |
-
|
352 |
-
def get_model_device(model):
|
353 |
-
if model is None:
|
354 |
-
return torch.device("cpu")
|
355 |
-
for p in model.parameters():
|
356 |
-
return p.device
|
357 |
-
return torch.device("cpu")
|
358 |
-
|
359 |
-
def clean_text(text):
|
360 |
-
return text.replace("</s>", "")
|
361 |
-
|
362 |
-
@torch.no_grad()
|
363 |
-
def response_report_generation(pil_image_1, pil_image_2):
|
364 |
-
"""Structured chest-X-ray report (streaming)."""
|
365 |
-
if not CHEXAGENT_AVAILABLE:
|
366 |
-
yield "CheXagent is not available. Please check installation."
|
367 |
-
return
|
368 |
-
|
369 |
-
streamer = TextIteratorStreamer(chex_tok, skip_prompt=True, skip_special_tokens=True)
|
370 |
-
paths = []
|
371 |
-
for im in [pil_image_1, pil_image_2]:
|
372 |
-
if im is None:
|
373 |
-
continue
|
374 |
-
with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tfile:
|
375 |
-
im.save(tfile.name)
|
376 |
-
paths.append(tfile.name)
|
377 |
-
|
378 |
-
if not paths:
|
379 |
-
yield "Please upload at least one image."
|
380 |
-
return
|
381 |
-
|
382 |
-
device = get_model_device(chex_model)
|
383 |
-
anatomies = [
|
384 |
-
"View",
|
385 |
-
"Airway",
|
386 |
-
"Breathing",
|
387 |
-
"Cardiac",
|
388 |
-
"Diaphragm",
|
389 |
-
"Everything else (e.g., mediastinal contours, bones, soft tissues, tubes, valves, pacemakers)",
|
390 |
-
]
|
391 |
-
prompts = [
|
392 |
-
"Determine the view of this CXR",
|
393 |
-
*[
|
394 |
-
f'Provide a detailed description of "{a}" in the chest X-ray'
|
395 |
-
for a in anatomies[1:]
|
396 |
-
],
|
397 |
-
]
|
398 |
-
|
399 |
-
findings = ""
|
400 |
-
partial = "## Generating Findings (step-by-step):\n\n"
|
401 |
-
for idx, (anat, prompt) in enumerate(zip(anatomies, prompts)):
|
402 |
-
query = chex_tok.from_list_format(
|
403 |
-
[*[{"image": p} for p in paths], {"text": prompt}]
|
404 |
-
)
|
405 |
-
conv = [
|
406 |
-
{"from": "system", "value": "You are a helpful assistant."},
|
407 |
-
{"from": "human", "value": query},
|
408 |
-
]
|
409 |
-
inp = chex_tok.apply_chat_template(
|
410 |
-
conv, add_generation_prompt=True, return_tensors="pt"
|
411 |
-
).to(device)
|
412 |
-
generate_kwargs = dict(
|
413 |
-
input_ids=inp,
|
414 |
-
max_new_tokens=512,
|
415 |
-
do_sample=False,
|
416 |
-
num_beams=1,
|
417 |
-
streamer=streamer,
|
418 |
-
)
|
419 |
-
Thread(target=chex_model.generate, kwargs=generate_kwargs).start()
|
420 |
-
partial += f"**Step {idx}: {anat}...**\n\n"
|
421 |
-
for tok in streamer:
|
422 |
-
if idx:
|
423 |
-
findings += tok
|
424 |
-
partial += tok
|
425 |
-
yield clean_text(partial)
|
426 |
-
partial += "\n\n"
|
427 |
-
findings += " "
|
428 |
-
findings = findings.strip()
|
429 |
-
|
430 |
-
# Impression
|
431 |
-
partial += "## Generating Impression\n\n"
|
432 |
-
prompt = f"Write the Impression section for the following Findings: {findings}"
|
433 |
-
conv = [
|
434 |
-
{"from": "system", "value": "You are a helpful assistant."},
|
435 |
-
{"from": "human", "value": chex_tok.from_list_format([{"text": prompt}])},
|
436 |
-
]
|
437 |
-
inp = chex_tok.apply_chat_template(
|
438 |
-
conv, add_generation_prompt=True, return_tensors="pt"
|
439 |
-
).to(device)
|
440 |
-
Thread(
|
441 |
-
target=chex_model.generate,
|
442 |
-
kwargs=dict(
|
443 |
-
input_ids=inp,
|
444 |
-
do_sample=False,
|
445 |
-
num_beams=1,
|
446 |
-
max_new_tokens=512,
|
447 |
-
streamer=streamer,
|
448 |
-
),
|
449 |
-
).start()
|
450 |
-
for tok in streamer:
|
451 |
-
partial += tok
|
452 |
-
yield clean_text(partial)
|
453 |
-
yield clean_text(partial)
|
454 |
-
|
455 |
-
@torch.no_grad()
|
456 |
-
def response_phrase_grounding(pil_image, prompt_text):
|
457 |
-
"""Very simple visual-grounding placeholder."""
|
458 |
-
if not CHEXAGENT_AVAILABLE:
|
459 |
-
return "CheXagent is not available. Please check installation.", None
|
460 |
-
|
461 |
-
if pil_image is None:
|
462 |
-
return "Please upload an image.", None
|
463 |
-
|
464 |
-
with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tfile:
|
465 |
-
pil_image.save(tfile.name)
|
466 |
-
img_path = tfile.name
|
467 |
-
|
468 |
-
device = get_model_device(chex_model)
|
469 |
-
query = chex_tok.from_list_format([{"image": img_path}, {"text": prompt_text}])
|
470 |
-
conv = [
|
471 |
-
{"from": "system", "value": "You are a helpful assistant."},
|
472 |
-
{"from": "human", "value": query},
|
473 |
-
]
|
474 |
-
inp = chex_tok.apply_chat_template(
|
475 |
-
conv, add_generation_prompt=True, return_tensors="pt"
|
476 |
-
).to(device)
|
477 |
-
out = chex_model.generate(
|
478 |
-
input_ids=inp, do_sample=False, num_beams=1, max_new_tokens=512
|
479 |
-
)
|
480 |
-
resp = clean_text(chex_tok.decode(out[0][inp.shape[1] :]))
|
481 |
-
|
482 |
-
# simple center box (placeholder)
|
483 |
-
w, h = pil_image.size
|
484 |
-
cx, cy, sz = w // 2, h // 2, min(w, h) // 4
|
485 |
-
draw = ImageDraw.Draw(pil_image)
|
486 |
-
draw.rectangle([(cx - sz, cy - sz), (cx + sz, cy + sz)], outline="red", width=3)
|
487 |
-
|
488 |
-
return resp, pil_image
|
489 |
-
|
490 |
-
# =============================================================================
|
491 |
-
# Gradio UI
|
492 |
-
# =============================================================================
|
493 |
-
def create_ui():
|
494 |
-
"""Create the Gradio interface."""
|
495 |
-
# Load Qwen model
|
496 |
-
try:
|
497 |
-
qwen_model, qwen_proc, qwen_dev = load_qwen_model_and_processor()
|
498 |
-
med_agent = MedicalVLMAgent(qwen_model, qwen_proc, qwen_dev)
|
499 |
-
qwen_available = True
|
500 |
-
except Exception as e:
|
501 |
-
print(f"Qwen model not available: {e}")
|
502 |
-
qwen_available = False
|
503 |
-
med_agent = None
|
504 |
-
|
505 |
-
with gr.Blocks(title="Medical AI Assistant") as demo:
|
506 |
-
gr.Markdown("# Combined Medical Q&A · SAM-2 Automatic Masking · CheXagent")
|
507 |
-
|
508 |
-
# Status information
|
509 |
-
with gr.Row():
|
510 |
-
gr.Markdown(f"""
|
511 |
-
**System Status:**
|
512 |
-
- Qwen VLM: {'✅ Available' if qwen_available else '❌ Not Available'}
|
513 |
-
- SAM-2: {'✅ Available' if SAM2_AVAILABLE else '❌ Not Available'}
|
514 |
-
- CheXagent: {'✅ Available' if CHEXAGENT_AVAILABLE else '❌ Not Available'}
|
515 |
-
""")
|
516 |
-
|
517 |
-
# Medical Q&A Tab
|
518 |
-
with gr.Tab("Medical Q&A"):
|
519 |
-
if qwen_available:
|
520 |
-
q_in = gr.Textbox(label="Question / description", lines=3)
|
521 |
-
q_img = gr.Image(label="Optional image", type="pil")
|
522 |
-
q_btn = gr.Button("Submit")
|
523 |
-
q_out = gr.Textbox(label="Answer")
|
524 |
-
q_btn.click(fn=med_agent.run, inputs=[q_in, q_img], outputs=q_out)
|
525 |
-
else:
|
526 |
-
gr.Markdown("❌ Medical Q&A is not available. Qwen model failed to load.")
|
527 |
-
|
528 |
-
# Segmentation Tab
|
529 |
-
with gr.Tab("Automatic masking"):
|
530 |
-
seg_img = gr.Image(label="Upload medical image", type="pil")
|
531 |
-
seg_btn = gr.Button("Run segmentation")
|
532 |
-
seg_out = gr.Image(label="Segmentation result", type="pil")
|
533 |
-
seg_status = gr.Textbox(label="Status", interactive=False)
|
534 |
-
|
535 |
-
if SAM2_AVAILABLE and _mask_generator is not None:
|
536 |
-
seg_btn.click(
|
537 |
-
fn=tumor_segmentation_interface,
|
538 |
-
inputs=seg_img,
|
539 |
-
outputs=[seg_out, seg_status],
|
540 |
-
)
|
541 |
-
else:
|
542 |
-
seg_btn.click(
|
543 |
-
fn=simple_segmentation_fallback,
|
544 |
-
inputs=seg_img,
|
545 |
-
outputs=[seg_out, seg_status],
|
546 |
-
)
|
547 |
-
|
548 |
-
# CheXagent Tabs
|
549 |
-
with gr.Tab("CheXagent – Structured report"):
|
550 |
-
if CHEXAGENT_AVAILABLE:
|
551 |
-
gr.Markdown("Upload one or two chest X-ray images; the report streams live.")
|
552 |
-
cx1 = gr.Image(label="Image 1", image_mode="L", type="pil")
|
553 |
-
cx2 = gr.Image(label="Image 2", image_mode="L", type="pil")
|
554 |
-
cx_report = gr.Markdown()
|
555 |
-
gr.Interface(
|
556 |
-
fn=response_report_generation,
|
557 |
-
inputs=[cx1, cx2],
|
558 |
-
outputs=cx_report,
|
559 |
-
live=True,
|
560 |
-
).render()
|
561 |
-
else:
|
562 |
-
gr.Markdown("❌ CheXagent structured report is not available.")
|
563 |
-
|
564 |
-
with gr.Tab("CheXagent – Visual grounding"):
|
565 |
-
if CHEXAGENT_AVAILABLE:
|
566 |
-
vg_img = gr.Image(image_mode="L", type="pil")
|
567 |
-
vg_prompt = gr.Textbox(value="Locate the highlighted finding:")
|
568 |
-
vg_text = gr.Markdown()
|
569 |
-
vg_out_img = gr.Image()
|
570 |
-
gr.Interface(
|
571 |
-
fn=response_phrase_grounding,
|
572 |
-
inputs=[vg_img, vg_prompt],
|
573 |
-
outputs=[vg_text, vg_out_img],
|
574 |
-
).render()
|
575 |
-
else:
|
576 |
-
gr.Markdown("❌ CheXagent visual grounding is not available.")
|
577 |
|
578 |
-
|
579 |
|
580 |
-
if __name__ == "__main__":
|
581 |
-
demo = create_ui()
|
582 |
-
demo.launch(server_name="0.0.0.0", server_port=7860, share=True)
|
|
|
1 |
#!/usr/bin/env python
|
|
|
|
|
2 |
"""
|
3 |
+
post_analyzer_enhanced.py · Enhanced Post Analysis Tool
|
4 |
+
=====================================================
|
5 |
|
6 |
+
Analyzes images of posts by running YOLOv8 inference, applying spatial layout rules,
|
7 |
+
computing a nuanced confidence score, and detecting anomalies ("afwijking").
|
8 |
+
Generates JSON reports for image directories and uploaded images.
|
9 |
+
Includes SAM-2 alias patch for Hugging Face compatibility.
|
|
|
|
|
10 |
"""
|
11 |
+
from __future__ import annotations
|
12 |
|
13 |
+
import argparse
|
14 |
+
import json
|
|
|
|
|
15 |
import sys
|
16 |
+
import os
|
|
|
17 |
import subprocess
|
18 |
+
import tempfile
|
19 |
+
from pathlib import Path
|
20 |
+
from typing import List, Union
|
21 |
+
from datetime import datetime
|
22 |
+
from urllib.parse import urlparse
|
|
|
23 |
|
24 |
+
import cv2
|
25 |
+
import yaml
|
|
|
|
|
26 |
import numpy as np
|
27 |
+
from dataclasses import dataclass
|
28 |
+
from ultralytics import YOLO
|
29 |
+
import requests
|
30 |
+
from PIL import Image
|
31 |
+
import io
|
32 |
+
|
33 |
+
# ───── Data Classes ──────────────────────────────────────────────────────────
|
34 |
+
@dataclass
|
35 |
+
class PostPart:
|
36 |
+
name: str
|
37 |
+
x: float # normalized center x
|
38 |
+
y: float # normalized center y
|
39 |
+
width: float
|
40 |
+
height: float
|
41 |
+
confidence: float = 1.0
|
42 |
+
|
43 |
+
@dataclass
|
44 |
+
class PostAnalysis:
|
45 |
+
image_path: Path
|
46 |
+
parts: List[PostPart]
|
47 |
+
anomalies: List[PostPart]
|
48 |
+
violations: List[str]
|
49 |
+
is_conform: bool
|
50 |
+
confidence_score: float
|
51 |
+
|
52 |
+
# ───── Configuration Load ────────────────────────────────────────────────────
|
53 |
+
def load_yaml_config(yaml_path: Path) -> dict:
|
54 |
+
if not yaml_path.exists():
|
55 |
+
sys.exit(f"Required {yaml_path} was not found – aborting.")
|
56 |
+
with yaml_path.open("r", encoding="utf-8") as fh:
|
57 |
+
data = yaml.safe_load(fh)
|
58 |
+
if "names" not in data:
|
59 |
+
sys.exit("'names' field missing in data.yaml – unable to continue.")
|
60 |
+
return {
|
61 |
+
"names": data["names"],
|
62 |
+
"class_to_name": {i: n for i, n in enumerate(data["names"])},
|
63 |
+
"name_to_class": {n: i for i, n in enumerate(data["names"])},
|
64 |
+
}
|
65 |
+
|
66 |
+
# ───── SAM-2 Alias Patch ─────────────────────────────────────────────────────
|
67 |
+
# Maps sam_2 package to sam2 namespace for correct imports
|
68 |
+
try:
|
69 |
+
import sam_2
|
70 |
+
import importlib
|
71 |
+
sys.modules['sam2'] = sam_2
|
72 |
+
for sub in ['build_sam','automatic_mask_generator','modeling.sam2_base']:
|
73 |
+
sys.modules[f'sam2.{sub}'] = importlib.import_module(f'sam_2.{sub}')
|
74 |
+
except ImportError:
|
75 |
+
pass
|
76 |
+
|
77 |
+
# ───── Dependency Checker & Installer (SAM-2) ─────────────────────────────────
|
78 |
+
def check_and_install_sam2() -> tuple[bool,str]:
|
79 |
try:
|
|
|
80 |
from sam2.build_sam import build_sam2
|
|
|
|
|
81 |
return True, "SAM-2 already available"
|
82 |
+
except ImportError:
|
83 |
+
# Clone if needed
|
84 |
+
if not os.path.exists("segment-anything-2"):
|
85 |
+
subprocess.run([
|
86 |
+
"git","clone",
|
87 |
+
"https://github.com/facebookresearch/segment-anything-2.git"
|
88 |
+
], check=True)
|
89 |
+
# Install editable
|
90 |
+
cwd = os.getcwd()
|
91 |
+
os.chdir("segment-anything-2")
|
92 |
+
subprocess.run([sys.executable, "-m", "pip", "install", "-e", "."], check=True)
|
93 |
+
os.chdir(cwd)
|
94 |
+
# Add to path and re-alias
|
95 |
+
path = os.path.abspath("segment-anything-2")
|
96 |
+
if path not in sys.path:
|
97 |
+
sys.path.insert(0, path)
|
98 |
try:
|
99 |
+
import sam_2, importlib
|
100 |
+
sys.modules['sam2'] = sam_2
|
101 |
+
for sub in ['build_sam','automatic_mask_generator','modeling.sam2_base']:
|
102 |
+
sys.modules[f'sam2.{sub}'] = importlib.import_module(f'sam_2.{sub}')
|
103 |
+
except ImportError:
|
104 |
+
return False, "SAM-2 import failed after install"
|
105 |
+
return True, "SAM-2 installed and aliased"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
106 |
|
|
|
107 |
SAM2_AVAILABLE, SAM2_STATUS = check_and_install_sam2()
|
108 |
print(f"SAM-2 Status: {SAM2_STATUS}")
|
|
|
|
|
|
|
|
|
109 |
if SAM2_AVAILABLE:
|
110 |
+
from sam2.build_sam import build_sam2
|
111 |
+
from sam2.automatic_mask_generator import SAM2AutomaticMaskGenerator
|
112 |
+
from sam2.modeling.sam2_base import SAM2Base
|
113 |
+
|
114 |
+
# ───── YOLO Inference ────────────────────────────────────────────────────────
|
115 |
+
def infer_parts(
|
116 |
+
img_path: Path,
|
117 |
+
model: YOLO,
|
118 |
+
class_info: dict,
|
119 |
+
) -> tuple[List[PostPart], List[PostPart]]:
|
120 |
+
results = model(str(img_path))
|
121 |
+
parts, anomalies = [], []
|
122 |
+
for det in results[0].boxes:
|
123 |
+
x, y, w, h = det.xywh[0].tolist()
|
124 |
+
cls_id = int(det.cls[0].item())
|
125 |
+
conf = float(det.conf[0].item())
|
126 |
+
name = class_info['class_to_name'].get(cls_id, f"unknown-{cls_id}")
|
127 |
+
part = PostPart(name, x, y, w, h, conf)
|
128 |
+
(anomalies if name=='afwijking' else parts).append(part)
|
129 |
+
return parts, anomalies
|
130 |
+
|
131 |
+
# ───── Spatial Validation ────────────────────────────────────────────────────
|
132 |
+
def check_position(part: PostPart, img_w: int, img_h: int) -> bool:
|
133 |
+
cx, cy = part.x*img_w, part.y*img_h
|
134 |
+
w_px, h_px = part.width*img_w, part.height*img_h
|
135 |
+
if part.name=='logo':
|
136 |
+
return (cx - w_px/2 >= 0.75*img_w) and (cy + h_px/2 <= 0.25*img_h)
|
137 |
+
return True
|
138 |
+
|
139 |
+
def validate_layout(parts: List[PostPart], image_shape: tuple[int,int]) -> List[str]:
|
140 |
+
img_h, img_w = image_shape
|
141 |
+
return [f"{p.name} out of expected zone" for p in parts if not check_position(p, img_w, img_h)]
|
142 |
+
|
143 |
+
# ───── Confidence Scoring ───────────────────────────────────────────────────
|
144 |
+
def compute_confidence(
|
145 |
+
parts: List[PostPart], anomalies: List[PostPart], violations: List[str]
|
146 |
+
) -> float:
|
147 |
+
base = sum(p.confidence for p in parts)/len(parts) if parts else 0.3
|
148 |
+
defect_penalty = min(0.1*len(anomalies), 0.5)
|
149 |
+
layout_penalty = min(0.05*len(violations), 0.3)
|
150 |
+
return max(0.0, base - defect_penalty - layout_penalty)
|
151 |
+
|
152 |
+
# ───── Core Analysis ────────────────────────────────────────────────────────
|
153 |
+
def analyze_post(
|
154 |
+
img_path: Path, model: YOLO, class_info: dict, quiet: bool=False
|
155 |
+
) -> PostAnalysis:
|
156 |
+
parts, anomalies = infer_parts(img_path, model, class_info)
|
157 |
+
img = cv2.imread(str(img_path))
|
158 |
+
if img is None: sys.exit(f"Failed to read image {img_path}")
|
159 |
+
violations = validate_layout(parts, img.shape[:2])
|
160 |
+
score = compute_confidence(parts, anomalies, violations)
|
161 |
+
conform = not anomalies and not violations
|
162 |
+
if not quiet:
|
163 |
+
status = 'CONFORM' if conform else 'NON-CONFORM'
|
164 |
+
print(f"{img_path.name}: {status} | parts={len(parts)}, anomalies={len(anomalies)}, violations={len(violations)} | score={score:.2f}")
|
165 |
+
return PostAnalysis(img_path, parts, anomalies, violations, conform, score)
|
166 |
+
|
167 |
+
# ───── Reporting ─────────────────────────────────────────────────────────────
|
168 |
+
def write_analysis_report(analyses: List[PostAnalysis], output_dir: Path) -> Path:
|
169 |
+
output_dir.mkdir(parents=True, exist_ok=True)
|
170 |
+
report = []
|
171 |
+
for a in analyses:
|
172 |
+
report.append({
|
173 |
+
'image': str(a.image_path), 'is_conform': a.is_conform,
|
174 |
+
'confidence_score': a.confidence_score, 'violations': a.violations,
|
175 |
+
'parts': [vars(p) for p in a.parts], 'anomalies': [vars(d) for d in a.anomalies]
|
176 |
+
})
|
177 |
+
fp = output_dir/'post_analysis.json'
|
178 |
+
with fp.open('w',encoding='utf-8') as f: json.dump(report,f,indent=2)
|
179 |
+
return fp
|
180 |
+
|
181 |
+
# ───── Image Download Helper ─────────────────────────────────────────────────
|
182 |
+
def download_image(url: str) -> Union[Path,None]:
|
183 |
try:
|
184 |
+
r = requests.get(url,timeout=10); r.raise_for_status()
|
185 |
+
parsed = urlparse(url)
|
186 |
+
ext = Path(parsed.path).suffix.lower() or '.jpg'
|
187 |
+
tmp = tempfile.NamedTemporaryFile(delete=False, suffix=ext)
|
188 |
+
tmp.write(r.content); tmp.close()
|
189 |
+
return Path(tmp.name)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
190 |
except Exception as e:
|
191 |
+
print(f"Download error for {url}: {e}"); return None
|
192 |
+
|
193 |
+
# ───── Process Uploaded Image ─────────────────────────────────────────────────
|
194 |
+
def process_uploaded_image(
|
195 |
+
image_data: Union[str,bytes,Path], model: YOLO, class_info: dict,
|
196 |
+
output_dir: Path, quiet: bool=False
|
197 |
+
) -> PostAnalysis:
|
198 |
+
tmp=None
|
|
|
|
|
199 |
try:
|
200 |
+
if isinstance(image_data,str) and image_data.startswith(('http://','https://')):
|
201 |
+
tmp = download_image(image_data); img_path=tmp or sys.exit()
|
202 |
+
elif isinstance(image_data,bytes):
|
203 |
+
img=Image.open(io.BytesIO(image_data)); fmt=img.format.lower(); ext=f".{fmt if fmt!='jpeg' else 'jpg'}"
|
204 |
+
tmp=tempfile.NamedTemporaryFile(delete=False,suffix=ext); tmp.write(image_data); tmp.close(); img_path=Path(tmp.name)
|
205 |
+
else:
|
206 |
+
img_path=Path(image_data);
|
207 |
+
if not img_path.exists(): sys.exit(f"File not found: {img_path}")
|
208 |
+
analysis = analyze_post(img_path, model, class_info, quiet)
|
209 |
+
out_fp = output_dir/f"analysis_{img_path.stem}.json"
|
210 |
+
with out_fp.open('w',encoding='utf-8') as f: json.dump({
|
211 |
+
'image':str(img_path),'is_conform':analysis.is_conform,
|
212 |
+
'confidence_score':analysis.confidence_score,'violations':analysis.violations,
|
213 |
+
'parts':[vars(p) for p in analysis.parts],'anomalies':[vars(d) for d in analysis.anomalies]
|
214 |
+
},f,indent=2)
|
215 |
+
return analysis
|
216 |
+
finally:
|
217 |
+
if tmp and Path(tmp.name).exists(): os.remove(tmp.name)
|
218 |
+
|
219 |
+
# ───── Process Directory & Uploads ───────────────────────────────────────────
|
220 |
+
def process_directory(images_dir: Path, output_dir: Path, data_yaml: Path, weights: str, quiet: bool=False):
|
221 |
+
ci=load_yaml_config(data_yaml); model=YOLO(weights)
|
222 |
+
imgs=[p for p in images_dir.iterdir() if p.suffix.lower() in ['.jpg','.jpeg','.png']]
|
223 |
+
if not imgs: sys.exit("No images found.")
|
224 |
+
output_dir.mkdir(parents=True,exist_ok=True)
|
225 |
+
analyses=[analyze_post(img,model,ci,quiet) for img in imgs]
|
226 |
+
rpt=write_analysis_report(analyses,output_dir)
|
227 |
+
print(f"Report written to {rpt}")
|
228 |
+
|
229 |
+
def process_uploaded_images(images: List[Union[str,bytes,Path]], output_dir: Path, data_yaml: Path, weights: str, quiet: bool=False):
|
230 |
+
ci=load_yaml_config(data_yaml); model=YOLO(weights); output_dir.mkdir(parents=True,exist_ok=True)
|
231 |
+
analyses=[]
|
232 |
+
for img in images:
|
233 |
+
try: analyses.append(process_uploaded_image(img,model,ci,output_dir,quiet))
|
234 |
+
except Exception as e: print(f"Error: {e}")
|
235 |
+
print(f"Processed {len(analyses)} uploads.")
|
236 |
+
return analyses
|
237 |
+
|
238 |
+
# ───── CLI Entrypoint ───────────────────────────────────────────────────────
|
239 |
+
def main(argv=None):
|
240 |
+
p=argparse.ArgumentParser(description="Enhanced post analysis tool")
|
241 |
+
p.add_argument("--images",type=Path,help="Directory of images")
|
242 |
+
p.add_argument("--upload",nargs="+",help="URLs, paths, or bytes to analyze")
|
243 |
+
p.add_argument("--output",type=Path,default="post_analysis_results")
|
244 |
+
p.add_argument("--data",type=Path,default="data.yaml")
|
245 |
+
p.add_argument("--weights",type=str,default="yolov8n.pt")
|
246 |
+
p.add_argument("-q","--quiet",action="store_true")
|
247 |
+
args=p.parse_args(argv)
|
248 |
+
if args.upload:
|
249 |
+
process_uploaded_images(args.upload,args.output,args.data,args.weights,args.quiet)
|
250 |
+
elif args.images:
|
251 |
+
process_directory(args.images,args.output,args.data,args.weights,args.quiet)
|
|
|
|
|
|
|
252 |
else:
|
253 |
+
p.error("Specify --images or --upload")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
254 |
|
255 |
+
if __name__ == "__main__": main()
|
256 |
|
|
|
|
|
|