from typing import List import requests import torch import base64 import numpy as np from PIL import Image from dotenv import load_dotenv from google import genai from google.genai import types import os load_dotenv() OPENAI_KEY = os.environ["OPENAI_API_KEY"] def encode_image_base64(path): with open(path, "rb") as f: return base64.b64encode(f.read()).decode("utf-8") def prepare_and_send_to_openai_api( png_paths, prompt, api_url="https://api.openai.com/v1/chat/completions", model="gpt-4o", ): headers = { "Authorization": f"Bearer {OPENAI_KEY}", "Content-Type": "application/json", } message_content = [{"type": "text", "text": prompt}] for p in png_paths: b64 = encode_image_base64(p) message_content.append( {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64}"}} ) payload = { "model": model, "messages": [{"role": "user", "content": message_content}], # "max_tokens": 1000, "temperature": 0.0, } # print(">>> sending msg") resp = requests.post(api_url, headers=headers, json=payload) if resp.status_code != 200: print("API error:", resp.status_code, resp.text) return None result = resp.json() return result["choices"][0]["message"]["content"] class GPTSliceVQATool: def run(self, png_paths, prompt): if isinstance(png_paths, str): png_paths = [png_paths] for image_path in png_paths: if image_path.endswith(".npy"): img = np.load(image_path) img = (img - np.min(img)) / (np.max(img) - np.min(img)) im = Image.fromarray((img * 255).astype(np.uint8)) image_path = image_path.replace(".npy", ".png") im.save(image_path) elif not image_path.endswith(".png"): return { "meta": None, "outputs": "ERROR: all input images must be a PNG or .npy file.", } png_paths = [ p if p.endswith(".png") else p.replace(".npy", ".png") for p in png_paths ] result = prepare_and_send_to_openai_api(png_paths=png_paths, prompt=prompt) return {"meta": None, "outputs": result} class GeminiSliceVQATool: def __init__(self, model="gemini-3-pro-preview"): self.client = genai.Client(api_key=os.getenv("GENAI_API_KEY")) self.model = model def run(self, png_paths, prompt): if isinstance(png_paths, str): png_paths = [png_paths] png_paths = [p.replace("'", "") for p in png_paths] for image_path in png_paths: if image_path.endswith(".npy"): img = np.load(image_path) img = (img - np.min(img)) / (np.max(img) - np.min(img)) im = Image.fromarray((img * 255).astype(np.uint8)) image_path = image_path.replace(".npy", ".png") im.save(image_path) elif not image_path.endswith(".png"): return { "meta": None, "outputs": f"ERROR: all input images must be a PNG or .npy file. Got <{image_path}>.", } png_paths = [ p if p.endswith(".png") else p.replace(".npy", ".png") for p in png_paths ] result = self.gemini_api_call(png_paths=png_paths, prompt=prompt) return {"meta": None, "outputs": result} def gemini_api_call(self, png_paths: List[str], prompt: str) -> str: inputs = [] for png_path in png_paths: with open(png_path, "rb") as f: image_bytes = f.read() inputs.append( types.Part.from_bytes( data=image_bytes, mime_type="image/png", ) ) inputs.append(prompt) response = self.client.models.generate_content( model=self.model, contents=inputs ) return response.text class vLLMSliceVQATool: def __init__(self): from vllm import LLM, SamplingParams os.environ["VLLM_USE_V1"] = "1" self.local_model = LLM( # model="Qwen/Qwen3-VL-30B-A3B-Instruct", # google/medgemma-1.5-4b-it model="google/gemma-3-27b-it", # "Qwen/Qwen3-VL-32B-Instruct-FP8", google/gemma-3-27b-it "OpenGVLab/InternVL2_5-4B" tensor_parallel_size=2, max_model_len=24000, gpu_memory_utilization=0.90, enable_chunked_prefill=True, max_num_batched_tokens=4096 * 8, # enforce_eager=True, # disable_custom_all_reduce=True, max_num_seqs=12, # 4 works trust_remote_code=True, ) self.sampling_params = SamplingParams(temperature=0.00, max_tokens=6000) def run_alternative(self, png_paths: List[str], prompt: str) -> dict: if isinstance(png_paths, str): png_paths = [png_paths] content = [] instruction = "You are a clinical expert analyzing a several chest CT slices. Please review the slices provided below carefully." content.append({"type": "text", "text": instruction}) for i, image_path in enumerate(png_paths, 1): content.append({"type": "text", "text": f"SLICE {i}"}) image_path = image_path.replace("'", "") if image_path.endswith(".npy"): img = np.load(image_path) img = (img - np.min(img)) / (np.max(img) - np.min(img)) content.append( { "type": "image_pil", "image_pil": Image.fromarray((img * 255).astype(np.uint8)), } ) elif image_path.endswith(".png"): content.append( {"type": "image_pil", "image_pil": Image.open(image_path)} ) elif image_path.endswith(".nii") or image_path.endswith(".nii.gz"): return { "meta": None, "outputs": "ERROR: CT volumes in NIfTI format are not supported. Select slices first.", } else: return { "meta": None, "outputs": "ERROR: all input images must be a PNG or .npy file.", } content.append({"type": "text", "text": prompt}) content.append( { "type": "text", "text": "Your response should be concise and focus on the main findings, ideally one paragraph only. Do NOT add disclaimer statements.", } ) with torch.no_grad(): outputs = self.local_model.chat( [{"role": "user", "content": content}], sampling_params=self.sampling_params, ) result = outputs[0].outputs[0].text return {"meta": None, "outputs": result} if __name__ == "__main__": from fastmcp import FastMCP from tool_configs import args_tools args = args_tools() mcp = FastMCP("see", stateless_http=False) slice_vqa_tool_instance = vLLMSliceVQATool() @mcp.tool() async def slice_vqa_tool(image_paths: List[str], question: str) -> dict: return slice_vqa_tool_instance.run_alternative( png_paths=image_paths, prompt=question ) mcp.run(transport="http", host=args.host, port=args.port)