""" app_core/visualization.py Convert model outputs to display-ready images for Gradio. All functions return uint8 numpy arrays [H, W, 3] or PIL Images. Gradio Image accepts both. """ import numpy as np from PIL import Image, ImageDraw, ImageFont import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt import matplotlib.cm as cm from typing import Optional def mask_to_rgb( binary_mask: np.ndarray, color: tuple = (45, 180, 150), # teal accent ) -> np.ndarray: """ Convert binary float mask [H,W] to coloured uint8 [H,W,3]. Positive pixels are teal, background is near-black. """ h, w = binary_mask.shape canvas = np.zeros((h, w, 3), dtype=np.uint8) canvas[binary_mask > 0.5] = color canvas[binary_mask <= 0.5] = (15, 20, 25) return canvas def prob_to_heatmap(prob_map: np.ndarray) -> np.ndarray: """ Convert probability map [H,W] to coloured heatmap uint8 [H,W,3]. Uses 'viridis' colormap for a clean academic look. """ normed = np.clip(prob_map, 0, 1) colored = (cm.viridis(normed)[:, :, :3] * 255).astype(np.uint8) return colored def make_overlay( img_np: np.ndarray, prob_map: np.ndarray, alpha: float = 0.45, threshold: float = 0.5, ) -> np.ndarray: """ Overlay probability map on CT image with a teal/hot colourmap. Args: img_np : float32 [H,W] normalised CT image prob_map : float32 [H,W] sigmoid probabilities alpha : overlay opacity threshold: draw contour at this level Returns: uint8 [H,W,3] """ # Base CT as RGB ct_rgb = np.stack([img_np, img_np, img_np], axis=2) ct_rgb = (ct_rgb * 255).clip(0, 255).astype(np.uint8) # Coloured probability overlay heat = (cm.plasma(np.clip(prob_map, 0, 1))[:, :, :3] * 255).astype(np.uint8) # Blend only where probability is meaningful mask_a = (prob_map > 0.2).astype(np.float32) blend = (ct_rgb.astype(np.float32) * (1 - alpha * mask_a[:,:,None]) + heat.astype(np.float32) * (alpha * mask_a[:,:,None])) result = blend.clip(0, 255).astype(np.uint8) # Draw contour at threshold using matplotlib try: fig, ax = plt.subplots(1, 1, figsize=(256/96, 256/96), dpi=96) ax.imshow(result) ax.contour(prob_map, levels=[threshold], colors=["#00e5cc"], linewidths=1.5) ax.axis("off") fig.tight_layout(pad=0) # Render to numpy array fig.canvas.draw() w_px, h_px = fig.canvas.get_width_height() buf = np.frombuffer(fig.canvas.tostring_rgb(), dtype=np.uint8) buf = buf.reshape(h_px, w_px, 3) plt.close(fig) # Resize back to 256×256 if needed if buf.shape[:2] != (256, 256): pil = Image.fromarray(buf).resize((256, 256), Image.BILINEAR) buf = np.array(pil) return buf except Exception: # Fallback if matplotlib rendering fails plt.close("all") return result def add_caption(img: np.ndarray, caption: str, font_size: int = 14) -> np.ndarray: """Add a text caption band at the bottom of an image.""" pil = Image.fromarray(img) band_h = font_size + 10 canvas = Image.new("RGB", (pil.width, pil.height + band_h), (18, 24, 32)) canvas.paste(pil, (0, 0)) draw = ImageDraw.Draw(canvas) draw.text((6, pil.height + 4), caption, fill=(180, 200, 210)) return np.array(canvas) def build_comparison_row( display_img: np.ndarray, results: dict, ) -> list: """ Build a list of (image, caption) tuples for Gradio Gallery. One entry per model: original, mask, overlay. """ gallery = [] gallery.append((display_img, "Input CT Slice")) for key, result in results.items(): if result["status"] == "ok": mask_img = result["binary_mask"] overlay_img = result["overlay"] else: h, w = display_img.shape[:2] mask_img = np.zeros((h, w, 3), dtype=np.uint8) overlay_img = display_img.copy() gallery.append((mask_img, f"{key} — Mask")) gallery.append((overlay_img, f"{key} — Overlay")) return gallery