File size: 10,241 Bytes
ee412eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import streamlit as st
from PIL import Image
import torch
import numpy as np
import os
from io import StringIO
import sys
import torch.nn as nn

# --- TorchDynamo Fix for Unsloth/MedGemma ---
import torch._dynamo
torch._dynamo.config.capture_scalar_outputs = True

# --- DEFINITIVE FIX FOR JIT COMPILER ERRORS ---
torch.compiler.disable()

# --- Dependency Handling ---
try:
    from monai.networks.nets import SwinUNETR
    import torchvision.transforms as T
    from unsloth import FastVisionModel
    from transformers import TextStreamer
    from s2wrapper import forward as multiscale_forward
except ImportError as e:
    st.error(f"A required library is not installed. Please install dependencies. Error: {e}")
    st.stop()

# --- Config and Model Definition ---
class Config:
    ORIGINAL_LABELS = [0,3,6,9,12,15,18,21,24,27,30,33,36,39,42,45,48,51,54,57,60]
    LABEL_MAP = {val: i for i, val in enumerate(ORIGINAL_LABELS)}
    NUM_CLASSES = len(ORIGINAL_LABELS)
    IMG_SIZE = (256, 256)
    FEATURE_SIZE = 48
    DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

class multiscaleSwinUNETR(nn.Module):
    def __init__(self, num_classes, scales=[1]):
        super().__init__()
        self.scales = scales
        self.num_classes = num_classes
        self.model = SwinUNETR(
            spatial_dims=2,
            in_channels=3,
            out_channels=num_classes,
            feature_size=Config.FEATURE_SIZE,
            drop_rate=0.0,
            attn_drop_rate=0.0,
            dropout_path_rate=0.0,
            use_checkpoint=True,
            use_v2=True
        )
        self.segmentation_head = nn.Sequential(
            nn.Conv2d(len(scales)*num_classes, num_classes, 3, padding=1),
            nn.BatchNorm2d(num_classes),
            nn.ReLU(inplace=True),
            nn.Conv2d(num_classes, num_classes, 1)
        )
    def forward(self, x):
        outs = multiscale_forward(self.model, x, scales=self.scales, output_shape="bchw")
        if isinstance(outs, (list, tuple)):
            normed = []
            for f in outs:
                f = f / (f.std(dim=(2, 3), keepdim=True) + 1e-6)
                normed.append(f)
            feats = torch.cat(normed, dim=1)
        elif isinstance(outs, torch.Tensor) and outs.dim() == 4:
            if len(self.scales) == 1:
                return outs
            feats = outs / (outs.std(dim=(2, 3), keepdim=True) + 1e-6)
        else:
            raise ValueError(f"Unexpected output shape/type from multiscale_forward: {type(outs)}, {getattr(outs,'shape',None)}")
        logits = self.segmentation_head(feats)
        return logits

# --- Model Loading ---
@st.cache_resource
def load_swinunetr_model():
    """Loads the multiscale SwinUNETR segmentation model."""
    model_path = 's2-swinunetr-weights.pth'
    if not os.path.exists(model_path):
        st.error(f"Segmentation model file not found at {model_path}")
        return None, None
    try:
        model = multiscaleSwinUNETR(num_classes=Config.NUM_CLASSES, scales=[1])
        model.load_state_dict(torch.load(model_path, map_location=Config.DEVICE))
        model.eval()
        return model, Config
    except Exception as e:
        st.error(f"Error loading segmentation model: {e}")
        return None, None

@st.cache_resource
def load_medgemma_model():
    """Loads the MedGemma vision-language model in eager mode."""
    try:
        model, processor = FastVisionModel.from_pretrained(
            "fiqqy/MedGemma-MM-OR-FT10",
            load_in_4bit=False,
            use_gradient_checkpointing="unsloth",
        )
        return model, processor
    except Exception as e:
        st.error(f"Error loading MedGemma model: {e}")
        return None, None

# --- Preprocessing ---
def preprocess_frames(frames, config):
    """Prepares image frames for the segmentation model."""
    transform = T.Compose([
        T.Resize(config.IMG_SIZE, antialias=True),
        T.ToTensor(),
        T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])
    tensors = [transform(frame.convert("RGB")) for frame in frames]
    batch = torch.stack(tensors)
    return batch

# --- Color Palette for Mask Visualization ---
def make_palette(num_classes):
    rng = np.random.default_rng(0)
    colors = rng.integers(0, 255, size=(num_classes, 3), dtype=np.uint8)
    colors[0] = np.array([0, 0, 0])
    return colors

# --- Inference ---
def run_segmentation(model, config, frames):
    """Runs segmentation on the uploaded frames and visualizes with a color palette."""
    st.write("Running segmentation...")
    batch = preprocess_frames(frames, config)
    device = config.DEVICE
    batch = batch.to(device)
    model = model.to(device)
    with torch.no_grad():
        logits = model(batch)
        preds = torch.argmax(logits, 1).cpu().numpy()
    mask = preds[0]
    st.write(f"Mask unique values: {np.unique(mask)}")
    palette = make_palette(config.NUM_CLASSES)
    color_mask = palette[mask]
    mask_img = Image.fromarray(color_mask.astype(np.uint8))
    return mask_img

# --- MedGemma Captioning ---
def run_captioning(medgemma_model, processor, frames, mask_img, instruction):
    """Runs MedGemma inference using 3 frames, 1 mask, and an instruction."""
    st.write("Preparing inputs for MedGemma...")
    images = [f.convert("RGB") for f in frames]
    mask_img = mask_img.convert("RGB")
    messages = [
        {"role": "user", "content": [
            {"type": "image"}, {"type": "image"}, {"type": "image"}, {"type": "image"},
            {"type": "text", "text": instruction},
        ]},
    ]
    input_text = processor.apply_chat_template(messages, add_generation_prompt=True)
    device = "cuda" if torch.cuda.is_available() else "cpu"
    all_images = images + [mask_img]
    inputs = processor(
        all_images, input_text, add_special_tokens=False, return_tensors="pt",
    ).to(device)

    text_streamer = TextStreamer(processor, skip_prompt=True)
    old_stdout = sys.stdout
    sys.stdout = captured_output = StringIO()

    st.write("Running MedGemma Analysis...")
    torch._dynamo.disable()
    medgemma_model.generate(
        **inputs, streamer=text_streamer, max_new_tokens=768,
        use_cache=True, temperature=1.0, top_p=0.95, top_k=64
    )

    sys.stdout = old_stdout
    result = captured_output.getvalue()
    return result

# --- Streamlit UI ---
def show():
    """Main function to render the Streamlit UI."""
    st.title("Surgical Scene Analysis System")
    st.write("A system to test surgical scene segmentation and captioning models.")

    st.header("1. Load Models")
    if "seg_model" not in st.session_state or "seg_config" not in st.session_state:
        st.session_state.seg_model, st.session_state.seg_config = None, None
    if st.button("Load Segmentation Model"):
        with st.spinner("Loading SwinUNETR..."):
            st.session_state.seg_model, st.session_state.seg_config = load_swinunetr_model()

    if st.session_state.seg_model is not None:
        st.success("Segmentation model is loaded.")
    else:
        st.warning("Segmentation model is not loaded.")

    if "medgemma_model" not in st.session_state:
        st.session_state.medgemma_model, st.session_state.processor = None, None
    if st.button("Load MedGemma Model"):
        with st.spinner("Loading MedGemma... This can take several minutes."):
            st.session_state.medgemma_model, st.session_state.processor = load_medgemma_model()

    if st.session_state.get("medgemma_model") and st.session_state.get("processor"):
        st.success("MedGemma model is loaded.")
    else:
        st.warning("MedGemma model is not loaded.")

    st.header("2. Upload Data & Generate Mask")
    st.subheader("Upload Three Sequential Surgical Video Frames")
    col1, col2, col3 = st.columns(3)
    uploaded_files = [
        col1.file_uploader("Upload Frame 1", type=["png", "jpg", "jpeg"], key="frame1"),
        col2.file_uploader("Upload Frame 2", type=["png", "jpg", "jpeg"], key="frame2"),
        col3.file_uploader("Upload Frame 3", type=["png", "jpg", "jpeg"], key="frame3")
    ]
    frames = [Image.open(f) for f in uploaded_files if f is not None]

    display_size = (256, 256)
    if "mask_img" not in st.session_state:
        st.session_state.mask_img = None

    if len(frames) == 3:
        st.success("All three frames have been uploaded successfully.")
        img_cols = st.columns(4)
        for i, frame in enumerate(frames):
            img_cols[i].image(frame.resize(display_size), caption=f"Frame {i+1}", use_container_width=True)

        if st.session_state.seg_model and st.session_state.seg_config and st.button("Run Segmentation"):
            with st.spinner("Generating segmentation mask..."):
                st.session_state.mask_img = run_segmentation(st.session_state.seg_model, st.session_state.seg_config, frames)

        if st.session_state.mask_img is not None:
            img_cols[3].image(st.session_state.mask_img.resize(display_size), caption="Segmentation Mask", use_container_width=True)
    else:
        st.info("Please upload all three frames to proceed.")

    st.header("3. Generate Scene Analysis")
    instruction_prompt = st.text_area(
        "Enter your custom instruction prompt:",
        "Provide a detailed summary of the surgical action, noting the instruments used and their interactions."
    )

    can_run_analysis = (
        st.session_state.get("medgemma_model") is not None and
        len(frames) == 3 and
        st.session_state.get("mask_img") is not None and
        bool(instruction_prompt)
    )

    if st.button("Run Analysis", disabled=not can_run_analysis):
        with st.spinner("Running MedGemma analysis... This may take a moment."):
            result = run_captioning(
                st.session_state.medgemma_model, st.session_state.processor,
                frames, st.session_state.mask_img, instruction_prompt
            )
            st.subheader("Analysis Result")
            st.write(result)

    if not can_run_analysis:
        st.warning("Please ensure the MedGemma model is loaded, three frames are uploaded, segmentation is complete, and a prompt is provided.")

if __name__ == "__main__":
    show()