import os import shutil import json import torch import torch.nn as nn import torchvision import numpy as np import scipy.ndimage as ndimage from skimage.measure import label, regionprops import gradio as gr import nibabel as nib from monai.transforms import ( Compose, LoadImaged, EnsureChannelFirstd, Resized, ScaleIntensityd, ToTensord, Lambdad, Orientationd, Spacingd, ScaleIntensityRangePercentilesd ) # ========================================== # 1. MODEL ARCHITECTURE # ========================================== def get_pretrained_gatekeeper(num_classes): model = torchvision.models.video.r3d_18(weights=None) original_stem = model.stem[0] new_stem = nn.Conv3d(1, original_stem.out_channels, kernel_size=original_stem.kernel_size, stride=original_stem.stride, padding=original_stem.padding, bias=False) model.stem[0] = new_stem model.fc = nn.Linear(model.fc.in_features, num_classes) return model # ========================================== # 2. INTELLIGENT AGENT SERVICE # ========================================== class GatekeeperService: def __init__(self, model_path): self.device = torch.device("cpu") self.classes = ["normal", "tumor", "ischemic"] print(f"--- 🔄 Initializing Gatekeeper Agent ---") self.model = get_pretrained_gatekeeper(num_classes=3) if os.path.exists(model_path): state_dict = torch.load(model_path, map_location=self.device) self.model.load_state_dict(state_dict) self.model.to(self.device) self.model.eval() print("✅ Model Loaded Successfully.") else: print(f"⚠️ Model file '{model_path}' not found.") # --- AI PIPELINE --- # Robust scaling for the AI to handle artifacts/skull self.transforms = Compose([ LoadImaged(keys=["image"]), EnsureChannelFirstd(keys=["image"]), Orientationd(keys=["image"], axcodes="RAS"), Spacingd(keys=["image"], pixdim=(1.5, 1.5, 1.5), mode=("bilinear")), Lambdad(keys=["image"], func=lambda x: x[0:1]), ScaleIntensityRangePercentilesd( keys=["image"], lower=1, upper=99, b_min=0.0, b_max=1.0, clip=True ), Resized(keys=["image"], spatial_size=(96, 96, 96)), ToTensord(keys=["image"]) ]) def identify_sequences(self, file_paths): flair_path = None swi_path = None debug_log = {} if not file_paths: return None, None, {} for path in file_paths: filename = os.path.basename(path).lower() debug_log[filename] = "Ignored" if "flair" in filename or "t2_flair" in filename: flair_path = path debug_log[filename] = "Matched: FLAIR" elif "swi" in filename or "gre" in filename or "hemo" in filename: swi_path = path debug_log[filename] = "Matched: SWI" return flair_path, swi_path, debug_log def physics_check_hemorrhage(self, swi_path): # (Physics logic remains unchanged as it was working correctly) if not swi_path: return False, 0.0, "N/A" try: img_obj = nib.load(swi_path) vol = img_obj.get_fdata() zooms = img_obj.header.get_zooms() voxel_vol = zooms[0] * zooms[1] * zooms[2] threshold = np.mean(vol) * 0.1 brain_mask = vol > threshold brain_center = ndimage.center_of_mass(brain_mask) brain_pixels = vol[brain_mask] if len(brain_pixels) > 0: p99 = np.percentile(brain_pixels, 99) vol = np.clip(vol / p99, 0, 1) dark_mask = (vol < 0.05) & brain_mask labeled_vol = label(dark_mask) regions = regionprops(labeled_vol) largest_bleed = 0.0 bleed_type = "Clean" for region in regions: size_mm3 = region.area * voxel_vol if 2000 < size_mm3 < 300000: centroid = region.centroid dist = np.linalg.norm(np.array(centroid) - np.array(brain_center)) min_r, min_c, min_d, max_r, max_c, max_d = region.bbox dims = sorted([max_r-min_r, max_c-min_c, max_d-min_d]) elongation = dims[2] / (dims[0] + 1e-5) if elongation < 3.0: if size_mm3 > largest_bleed: largest_bleed = size_mm3 bleed_type = f"Deep Lesion (Vol: {size_mm3/1000:.1f}mL)" if largest_bleed > 0: return True, largest_bleed, bleed_type return False, 0.0, "Artifacts Filtered" except Exception as e: print(f"Physics Error: {e}") return False, 0.0, "Error" def process_batch(self, file_list): flair_path, swi_path, file_log = self.identify_sequences(file_list) response = { "status": "PROCESSING", "action": "WAIT", "findings": [], "debug_files": file_log } if not flair_path: response["status"] = "FAILED" response["findings"].append("Error: No FLAIR sequence found.") return response # --- STEP 1: SAFETY GATE (Physics) --- if swi_path: is_bleed, size, shape = self.physics_check_hemorrhage(swi_path) if is_bleed: response["status"] = "COMPLETED" response["action"] = "CONSULT_REPORT" response["findings"].append(f"SAFETY STOP: Suspected Hemorrhage ({shape}) on SWI.") return response else: response["findings"].append("Safety Check: No active hemorrhage detected.") # --- STEP 2: TRIAGE (AI) --- try: input_dict = [{"image": flair_path}] tensor_data = self.transforms(input_dict)[0]["image"].unsqueeze(0).to(self.device) with torch.no_grad(): output = self.model(tensor_data) probs = torch.softmax(output, dim=1) # Probabilities # 0: Normal, 1: Tumor, 2: Ischemic p_tumor = probs[0, 1].item() p_ischemic = probs[0, 2].item() pred_idx = torch.argmax(probs, dim=1).item() diagnosis = self.classes[pred_idx] # --- RESTORED FILTERING LOGIC --- # 1. Clear Stroke -> Go if p_ischemic > 0.35: response["status"] = "COMPLETED" response["action"] = "ACTIVATE_VISION_AGENT" response["findings"].append(f"Triage: Ischemic Stroke Candidate (Confidence {p_ischemic:.2%}).") # 2. Clear Tumor -> Stop (THIS IS THE FIX) # If the model is >80% sure it's a Tumor, we trust it and stop the pipeline. elif diagnosis == "tumor" and p_tumor > 0.80: response["status"] = "COMPLETED" response["action"] = "CONSULT_REPORT" # <--- NOW IT STOPS response["findings"].append(f"Triage: High Confidence Tumor ({p_tumor:.2%}). Oncology referral suggested. Vision Agent Skipped.") # 3. Ambiguous Tumor -> Go (Safety) # If it's a Tumor but confidence is low (60%), we double check with Vision Agent. elif diagnosis == "tumor": response["status"] = "COMPLETED" response["action"] = "ACTIVATE_VISION_AGENT" response["findings"].append(f"Triage: Ambiguous Mass ({p_tumor:.2%}). Routing to Vision Agent to verify if Ischemic.") # 4. Normal/Other -> Stop else: response["status"] = "COMPLETED" response["action"] = "CONSULT_REPORT" response["findings"].append("Triage: No acute findings.") except Exception as e: response["status"] = "FAILED" response["findings"].append(f"AI Error: {str(e)}") return response # ========================================== # 3. INITIALIZATION & UI # ========================================== MODEL_FILE = "gatekeeper_hybrid_trained.pth" agent = GatekeeperService(MODEL_FILE) def predict_mri_protocol(file_list): return agent.process_batch(file_list) with gr.Blocks() as demo: gr.Markdown("# 🧠 Neuro-Triage: Gatekeeper Node") with gr.Row(): with gr.Column(): file_input = gr.File(label="Upload MRI Protocol", file_count="multiple", type="filepath") submit_btn = gr.Button("Triaging Patient", variant="primary") with gr.Column(): json_output = gr.JSON(label="Orchestrator Routing Instructions") submit_btn.click(predict_mri_protocol, inputs=[file_input], outputs=[json_output]) if __name__ == "__main__": demo.launch()