config.py

"""
Central configuration module for the Thyroid Nodule Analysis Pipeline.
All constants, model paths, feature definitions, and clinical lookup tables
are defined here.
"""

import os
import torch

# Environment detection
ON_HF_SPACES = os.environ.get("SPACE_ID") is not None

# Paths
BASE       = os.path.dirname(os.path.abspath(__file__))
YOLO_W     = os.path.join(BASE, "models", "yolo_finetuned_thyroidxl.pt")
UNET_W     = os.path.join(BASE, "models", "best_unet_finetuned.pth")
RESNET_W   = os.path.join(BASE, "models", "resnet50_finetuned_thyroidxl.pth")
TIRADS_W   = os.path.join(BASE, "models", "best_model_expB_RandomForest.pkl")
TRAIN_CSV  = os.path.join(BASE, "models", "train_features_patient_level_unetmask.csv")

# Device  (auto-detected: GPU if available, CPU otherwise)
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Model hyper-parameters
YOLO_CONF       = 0.35          # YOLO confidence threshold
YOLO_IOU_THRESH = 0.50          # YOLO IoU threshold for NMS
YOLO_IMGSZ      = 640           # YOLO inference image size

UNET_IMG_SIZE   = 384           # UNet++ input resolution (H = W)
RESNET_IMG_SIZE = 224           # ResNet50 input resolution (H = W)

IMAGENET_MEAN   = [0.485, 0.456, 0.406]   # ImageNet normalisation mean
IMAGENET_STD    = [0.229, 0.224, 0.225]   # ImageNet normalisation std

# Radiomic feature columns  (25 features - match Exp B training order)
FEAT_COLS = [
    # Morphological / shape
    "area_px", "aspect_ratio", "solidity", "circularity", "fractal_dim",
    "cv_radial", "elongation",
    # Intensity / echogenicity
    "mean_nodule", "std_nodule", "skewness", "kurt", "entropy_shannon",
    "cv_intensity", "rer", "solid_fraction",
    # GLCM texture
    "glcm_contrast", "glcm_asm", "glcm_idm", "glcm_correlation",
    "glcm_dissimilarity", "glcm_entropy",
    # Echogenic foci (calcification proxy)
    "calcif_density", "n_blobs", "mean_blob_size", "peripheral_ratio",
]

# Top-10 features selected from Exp B Random Forest feature importances
DISPLAY_FEATURES = [
    "aspect_ratio", "rer", "std_nodule", "cv_intensity",
    "skewness", "elongation", "solid_fraction",
    "entropy_shannon", "peripheral_ratio", "glcm_correlation",
]

FEATURE_LABELS = {
    "aspect_ratio"    : "Aspect Ratio",
    "rer"             : "Relative Echogenicity Ratio",
    "std_nodule"      : "Intensity Std Dev",
    "cv_intensity"    : "CV Intensity",
    "skewness"        : "Skewness",
    "elongation"      : "Elongation",
    "solid_fraction"  : "Solid Fraction",
    "entropy_shannon" : "Shannon Entropy",
    "peripheral_ratio": "Peripheral Calcif. Ratio",
    "glcm_correlation": "GLCM Correlation",
}

FEATURE_CATEGORY = {
    "aspect_ratio"    : "Shape",
    "rer"             : "Echogenicity",
    "std_nodule"      : "Echogenicity",
    "cv_intensity"    : "Echogenicity",
    "skewness"        : "Intensity",
    "elongation"      : "Shape",
    "solid_fraction"  : "Composition",
    "entropy_shannon" : "Intensity",
    "peripheral_ratio": "Echogenic Foci",
    "glcm_correlation": "Texture (GLCM)",
}

# Patient-level TI-RADS aggregation strategy
# MAX - higher value = more suspicious  ->  keep worst-case across views
TIRADS_AGG_MAX = [
    "aspect_ratio", "fractal_dim", "cv_radial", "elongation",
    "calcif_density", "n_blobs", "mean_blob_size",
    "peripheral_ratio", "cv_intensity", "skewness", "kurt",
]
# MIN - lower value = more suspicious -> keep worst-case across views
TIRADS_AGG_MIN = ["solidity", "circularity", "rer"]
# MEAN - remaining features (intensity statistics, GLCM)

# ACR TI-RADS clinical lookup tables
TIRADS_RISK = {
    1: "Benign",
    2: "Not suspicious",
    3: "Mildly suspicious",
    4: "Moderately suspicious",
    5: "Highly suspicious",
}

TIRADS_FNAB = {
    1: "Not recommended",
    2: "Not recommended",
    3: "Recommended if ≥ 2.5 cm",
    4: "Recommended if ≥ 1.5 cm",
    5: "Recommended if ≥ 1.0 cm",
}

# Accepted image formats for upload validation
ACCEPTED_EXTENSIONS = {".png", ".jpg", ".jpeg"}
ACCEPTED_MIME_TYPES = {"image/png", "image/jpeg"}