File size: 14,433 Bytes
7376e8f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a204636
7376e8f
0bbb9a2
7376e8f
 
 
 
 
 
 
 
 
 
c01165b
081aab9
7376e8f
e2def54
7376e8f
 
0bbb9a2
4678e3d
 
0bbb9a2
7376e8f
0bbb9a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7376e8f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0bbb9a2
10d9809
7376e8f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8cf8dba
7376e8f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0bbb9a2
 
 
 
 
 
 
 
 
44506cf
 
 
 
0bbb9a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7376e8f
 
 
0bbb9a2
7376e8f
f533365
7376e8f
f533365
7376e8f
 
 
 
 
28cf2c9
ecfbc3e
7376e8f
 
 
 
0bbb9a2
1489d40
0bbb9a2
28cf2c9
0bbb9a2
7376e8f
 
 
0bbb9a2
 
7376e8f
 
 
 
 
 
 
 
 
0bbb9a2
 
 
9108ad3
0bbb9a2
9108ad3
 
0bbb9a2
 
9108ad3
0bbb9a2
 
 
 
44506cf
0bbb9a2
7376e8f
0bbb9a2
7376e8f
 
 
a08e4df
12722f4
a08e4df
 
7376e8f
 
081aab9
12722f4
7376e8f
 
12722f4
7376e8f
6bcaa34
7376e8f
0bbb9a2
 
6bcaa34
e2def54
7376e8f
6bcaa34
0bbb9a2
 
7376e8f
 
cb1e133
6bcaa34
7376e8f
 
 
 
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
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
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
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
# -*- coding: utf-8 -*-
"""gradio_app_chestvision-PRO.ipynb

Automatically generated by Colab.

Original file is located at
    https://colab.research.google.com/drive/1gVrx5TyipNPvn8D7GaK0pNBCnLeYTAD_
"""

"""### Import dependencies"""

import torch
import torch.nn.functional as F
import torchvision
from torchvision import transforms, models, datasets
from torch import nn, optim
from torch.utils.data import DataLoader, Dataset
from tqdm import tqdm
from torch.utils.data import random_split
import pytorch_lightning as torch_light
from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint
import torchmetrics
from torchmetrics import Metric
import os
import shutil
import subprocess
import pandas as pd
from PIL import Image
import gradio
from functools import partial
from transformers import AutoModel, pipeline

"""### Initialize Containers"""

configs = {
    "IMAGE_SIZE":   (224, 224),    # Resize images to (W, H)
    "NUM_CHANNELS": 3,             # RGB images
    "NUM_CLASSES":  15,            # Number of output labels

    # ImageNet dataset normalization values (for pretrained backbones)
    "MEAN": (0.485, 0.456, 0.406),
    "STD":  (0.229, 0.224, 0.225),

    "DEFAULT_BACKBONE": "CheXFormer-small",
    "DEFAULT_VLM": "Lingshu-7B",

    "THRESHOLD": 0.2
    }

ViT_REGISTRY = {
    "CheXFormer-small": "m42-health/CXformer-small",
    # "CheXFormer-base": "m42-health/CXformer-base",
    "ViT-base-16": "google/vit-base-patch16-224"}

VLM_REGISTRY = {
    "MedMO": "MBZUAI/MedMO-8B",
    "Qwen3-VL-2B": "Qwen/Qwen3-VL-2B-Instruct",
    "Lingshu-7B": "lingshu-medical-mllm/Lingshu-7B",
    "MedGemma-4b": "google/medgemma-1.5-4b-it"}

VLM_SYSTEM_PROMPT = """ You are a medical imaging assistant specializing in chest radiography.

      A trained multi-label classifier analyzed a chest X-ray and made a prediction, including predicted medical condition(s) and their associated probabilities:

      Your task:
      1. Analyze the chest X-ray image to identify key features supporting the predicted condition(s).
      2. Do NOT introduce new diagnoses.
      3. Only explain radiographic findings that could support the listed prediction(s).
      4. Use cautious, uncertainty-aware language.
      5. If probability < 0.50, emphasize uncertainty.
      6. Do NOT contradict the classifier.

      Structure your answer as:

      Observed Radiographic Findings:
      ...

      How Chest X-ray Features Support the Predicted Conditions:
      ...
      """

ViT_MODEL_CACHE = {}
VLM_MODEL_CACHE = {}

"""### Define helper functions"""

# helper function for loading pre-trained model
# ===================================================================================================
class get_pretrained_model(nn.Module):
    def __init__(
        self,
        model_name: str,
        num_classes: int,
        num_layers_to_unfreeze: int = 0):
        super().__init__()

        print(f"Loading pretrained [{model_name}] model")

        self.backbone = AutoModel.from_pretrained(
            ViT_REGISTRY[model_name],
            # model_name,
            trust_remote_code=True)

        hidden_size = self.backbone.config.hidden_size

        # Freeze entire backbone first
        for param in self.backbone.parameters():
            param.requires_grad = False

        # Selectively unfreeze last N layers
        if num_layers_to_unfreeze > 0:
            self._unfreeze_last_n_layers(num_layers_to_unfreeze)

        # Single classification head
        self.classifier = nn.Sequential(
            nn.LayerNorm(hidden_size),
            nn.Dropout(0.4),
            nn.Linear(hidden_size, num_classes) )

    def forward(self, x):
        outputs = self.backbone(x)

        # Use CLS token
        img_embeddings = outputs.last_hidden_state[:, 0]

        logits = self.classifier(img_embeddings)
        return logits

    def _unfreeze_last_n_layers(self, n: int):
        if hasattr(self.backbone, "encoder"):
            encoder_layers = self.backbone.encoder.layer
        elif hasattr(self.backbone, "vision_model"):
            encoder_layers = self.backbone.vision_model.encoder.layer
        else:
            raise ValueError("Cannot find encoder layers in backbone.")

        total_layers = len(encoder_layers)
        n = min(n, total_layers)

        print(f"Unfreezing last {n} of {total_layers} transformer layers.")

        for layer in encoder_layers[-n:]:
            for param in layer.parameters():
                param.requires_grad = True


# helper function for preprocessing input images
# ===================================================================================================
preprocess_fxn = transforms.Compose(
      [transforms.Resize(size=configs["IMAGE_SIZE"][::-1]),
      transforms.ToTensor(),
      transforms.Normalize(configs["MEAN"], configs["STD"], inplace=True)])

# Map numeric outputs to string labels
labels_dict = {
    0: "Atelectasis",
    1: "Cardiomegaly",
    2: "Consolidation",
    3: "Edema",
    4: "Effusion",
    5: "Emphysema",
    6: "Fibrosis",
    7: "Hernia",
    8: "Infiltration",
    9: "Mass",
    10: "No finding",
    11: "Nodule",
    12: "Pleural_Thickening",
    13: "Pneumonia",
    14: "Pneumothorax"}

"""### Create torch lightning model (i.e., classifier) module"""

class modelModule(torch_light.LightningModule):
    def __init__(self, backbone_model_name, num_layers_to_unfreeze, num_classes=configs['NUM_CLASSES']):
        super().__init__()
        self.num_classes = num_classes
        self.backbone_model_name = backbone_model_name
        self.num_layers_to_unfreeze = num_layers_to_unfreeze

        # Load a pretrained backbone and replace its final layer
        self.model = get_pretrained_model(
            num_classes = self.num_classes,
            model_name  = self.backbone_model_name,
            num_layers_to_unfreeze = self.num_layers_to_unfreeze)

        # Binary classification loss operating on raw logits
        self.loss_function      = torch.nn.BCEWithLogitsLoss()

        self.accuracy_function  = torchmetrics.classification.MultilabelAccuracy(num_labels=self.num_classes, average="weighted", threshold=0.5)
        self.f1_score_function  = torchmetrics.classification.MultilabelF1Score(num_labels=self.num_classes, average="weighted", threshold=0.5)
        self.auroc_function     = torchmetrics.classification.MultilabelAUROC(num_labels=self.num_classes, average="weighted", thresholds=10)
        self.map_score_function = torchmetrics.classification.MultilabelAveragePrecision(num_labels=self.num_classes, average="weighted", thresholds=10)
        # average options: macro (simple average), micro (sum), weighted (weight by class size, then avg)
        # threshold: Threshold for transforming probability to binary (0,1) predictions. For some metrics (e.g., AUROC), represents the number of thresholds (evenly spaced b/n 0–1) the metric should be computed at (resulting array of values are the averaged to obtain the final score)

    def forward(self, x):
        # Forward pass through the backbone model
        return self.model(x)

    def _common_step(self, batch, batch_idx):
        """
        Shared logic for train / val / test steps.
        Computes loss and evaluation metrics.
        """
        x, y = batch

        # Compute model predictions ()
        y_logits = self.forward(x)
        y_prob    = torch.sigmoid(y_logits)

        # Compute metrics (expects logits + labels)
        loss     = self.loss_function(y_logits, y.float())

        # Compute mean loss over all classes
        # loss     = torchmetrics.aggregation.MeanMetric(self.loss_function(y_hat, y.float()), weight=X.shape[0])
        accuracy = self.accuracy_function(y_prob, y)
        f1_score = self.f1_score_function(y_prob, y)
        auroc    = self.auroc_function(y_prob, y)
        mAP      = self.map_score_function(y_prob, y) # mean average precision

        return loss, y_logits, y, accuracy, f1_score, auroc, mAP

    def training_step(self, batch, batch_idx):
        # Run shared step
        loss, y_logits, y, accuracy, f1_score, auroc, mAP = self._common_step(batch, batch_idx)

        # Log epoch-level training metrics
        self.log_dict(
            {"train_loss": loss, "train_accuracy": accuracy, "train_f1_score": f1_score, "train_auroc": auroc, "train_mAP": mAP},
            on_step=False, on_epoch=True, prog_bar=True)

        # Lightning expects the loss key for backprop
        return {"loss": loss}

    def validation_step(self, batch, batch_idx):
        # Run shared step
        loss, y_logits, y, accuracy, f1_score, auroc, mAP = self._common_step(batch, batch_idx)

        # Log validation metrics
        self.log_dict(
            {"val_loss": loss, "val_accuracy": accuracy,"val_f1_score": f1_score, "val_auroc": auroc, "val_mAP": mAP},
            on_step=False, on_epoch=True, prog_bar=True)

    def test_step(self, batch, batch_idx):
        # Run shared step
        loss, y_logits, y, accuracy, f1_score, auroc, mAP = self._common_step(batch, batch_idx)

        # Log test metrics
        self.log_dict(
            {"test_loss": loss, "test_accuracy": accuracy,"test_f1_score": f1_score, "test_auroc": auroc, "test_mAP": mAP},
            on_step=False, on_epoch=True, prog_bar=True)

    def predict_step(self, batch, batch_idx):
        """
        Prediction logic used by trainer.predict().
        Returns model outputs without computing loss.
        """
        x = batch if not isinstance(batch, (tuple, list)) else batch[0]
        logits = self.forward(x)

        # Convert logits to probabilities for inference
        probs = torch.sigmoid(logits)

        return probs

    def configure_optimizers(self):
        # Optimizer over all trainable parameters
        optimizer = optim.Adam(self.parameters(), lr=3e-5)
        return optimizer

"""### Create function for running inference (i.e., assistive medical diagnosis)"""

def generate_query(formatted_predictions):
    return f"""
      The predicted conditions and their corresponding probabilities are given by the following dictionary:

      {formatted_predictions}

      What features of the chest X-ray image support the predicted condition(s)?
      """

def predictionReportGenerator(vlm_model, img_pil, system_prompt, query_prompt):
    # image_ = Image.open(image_path).convert("RGB")
    image_ = img_pil.convert("RGB")

    messages = [
    {
        "role": "system",
        "content": [{"type": "text", "text": f"{system_prompt}"}]},
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image_},
            {"type": "text", "text": f"{query_prompt}"}]}]

    output = vlm_model(text=messages, max_new_tokens=350)
    prediction_explanation = output[0]["generated_text"][-1]["content"]
    
    return prediction_explanation


    
@torch.inference_mode()
def run_diagnosis(
    backbone_name,
    vlm_name,
    input_image,
    threshold,
    preprocess_fn=None,
    Idx2labels=None):

    # Preprocess
    x = preprocess_fn(input_image).unsqueeze(0)

    # Resolve backbone
    # ckpt_path = os.path.join(CKPT_ROOT, MODEL_REGISTRY[backbone_name])
    ckpt_path = os.path.join(CKPT_ROOT, f"{backbone_name}.ckpt")

    if not os.path.exists(ckpt_path):
        raise FileNotFoundError(f"Checkpoint not found: {ckpt_path}")

    # Load classification model (cache for speed)
    if backbone_name not in ViT_MODEL_CACHE:
      ViT_MODEL_CACHE[backbone_name] = modelModule.load_from_checkpoint(
        ckpt_path, backbone_model_name=backbone_name, num_layers_to_unfreeze = 2)
    model = ViT_MODEL_CACHE[backbone_name]

    model.eval()

    # device = 0 if torch.cuda.is_available() else -1

    # Forward
    logits = model(x)
    probs = torch.sigmoid(logits)[0].cpu().numpy()

    output_probs = {
        Idx2labels[i]: float(p) for i, p in enumerate(probs)
    }

    predicted_classes = [
        Idx2labels[i] for i, p in enumerate(probs) if p >= threshold]

    explanation_ = "No prediction was made."
    if predicted_classes:
        # Load model (cache for speed)
        if vlm_name not in VLM_MODEL_CACHE:
          VLM_MODEL_CACHE[vlm_name]     = pipeline(task  = "image-text-to-text",
                              model  = VLM_REGISTRY[vlm_name],
                              trust_remote_code = True)
        VLM_model = VLM_MODEL_CACHE[vlm_name]

        formatted_predictions = {label: output_probs[label] for label in predicted_classes}
        
        query_prompt = generate_query(formatted_predictions)
        explanation_ = predictionReportGenerator(vlm_model = VLM_model, img_pil = input_image, 
                                                 system_prompt = VLM_SYSTEM_PROMPT, query_prompt = query_prompt)

    return "\n".join(predicted_classes), explanation_, output_probs

"""### Gradio app"""

CKPT_ROOT = os.path.join(os.getcwd(), "Vision Transformers")

example_list_dir = os.path.join(os.getcwd(), "Curated test samples")
example_list_img_names = os.listdir(example_list_dir)

example_list = [
    [configs["DEFAULT_BACKBONE"], configs["DEFAULT_VLM"], os.path.join(example_list_dir, example_img)]
    for example_img in example_list_img_names[:8]
    if example_img.lower().endswith(".png")]


gradio_app = gradio.Interface(
    fn     = partial(run_diagnosis, preprocess_fn = preprocess_fxn, Idx2labels = labels_dict),

    inputs = [gradio.Dropdown(["CheXFormer-small", "ViT-base-16"], value="CheXFormer-small", label="Select Classification Model"),
              gradio.Dropdown(["MedGemma-4b", "MedMO", "Lingshu-7B", "Qwen3-VL-2B"], value="Lingshu-7B", label="Select Explanation Model"),
              gradio.Image(type="pil", label="Load chest-X-ray image here"),
              gradio.Slider(minimum = 0.1, maximum = 0.9, step = 0.05, value = 0.2, label = "Set Prediction Threshold")],

    outputs = [gradio.Textbox(label="Predicted Medical Condition(s)"),
               gradio.Textbox(label="Prediction Report"),
               gradio.Label(label="Predicted Probabilities", show_label=False)],

    examples       = example_list,
    cache_examples = False,
    title          = "ChestVision-PRO",
    description    = "Vision-Transformer solutions for assistive medical diagnosis with Vision-Language-based prediction justification",
    article        = "Author: C. Foli (02.2026) | Website: coming soon...")

gradio_app.launch()