Warn that the resize operator is part of the model; fix the usage snippet

84c698a verified 5 days ago

4.86 kB

license: mit
tags:
  - medical-imaging
  - chest-xray
  - pneumonia
  - densenet
  - onnx
  - image-classification
library_name: onnxruntime
pipeline_tag: image-classification

PneumoScan — Chest X-ray Pneumonia Triage

DenseNet121 model that flags likely pneumonia on frontal chest radiographs, built as a triage tool with a trustworthiness-first evaluation (external validation, calibration, uncertainty-aware abstention, subgroup fairness, Grad-CAM).

⚠️ Not a medical device. Research/education prototype only. Not for clinical use.

Live demo: https://pneumonia-detection-cnn-chi.vercel.app
Code + full write-up: https://github.com/decstar0610/pneumonia-detection-cnn
Model card (limitations): https://github.com/decstar0610/pneumonia-detection-cnn/blob/main/MODEL_CARD.md

Files in this repo

File	What it is
`backbone.onnx`	DenseNet121 feature extractor (image 224×224×3 → 7×7×1024), for onnxruntime
`gradcam_head.npz`	Classification-head weights (fc1/bn1/fc2/bn2/prob) — run in NumPy for prediction + Grad-CAM
`best_model.keras`	Original TensorFlow/Keras model (reference; the ONNX+NumPy path is validated identical)
`threshold.json`	Operating threshold (0.324) tuned on validation for sensitivity ≥ 0.92
`temperature.json`	Calibration temperature (T = 0.72)
`triage.json`	Calibrated threshold + abstention band for the 3-zone triage

Key results

Internal test: sensitivity 0.942, specificity 0.962, ROC-AUC 0.986.
External validation (RSNA, adults, n=3000): sensitivity holds (0.907) but specificity collapses (0.467) — the model was trained on pediatric data and false-alarms on unseen adult negatives. This is disclosed, not hidden.
PA-view sensitivity 0.615 (vs 0.985 AP) — a real subgroup weakness.

See the model card for full limitations.

⚠️ The resize operator is part of the model

Read this before writing your own preprocessing.

This model was trained and evaluated with tf.image.resize, which is bilinear without antialiasing. PIL.Image.resize — with its default resample or with BILINEAR — antialiases when downscaling. It is a different operator, and swapping it in measurably changes the model.

Measured over the full 586-image held-out test split:

Resize	Sensitivity	ROC-AUC	False negatives
`PIL.Image.resize(..., BILINEAR)`	0.9089	0.9830	39
`tf.image`-equivalent (correct)	0.9439	0.9863	24

That is 15 lost true positives out of 428 and 24 flipped decisions — from a substitution that looks harmless and produces per-image probability differences of ~1e-4 on confident cases. The damage is concentrated near the decision threshold, so a spot check of a few images will not reveal it.

The usage snippet below implements the correct resize.

Usage (onnxruntime + NumPy head)

import numpy as np, onnxruntime as ort
from PIL import Image

def resize_bilinear(arr, size=224):
    """Matches tf.image.resize(..., antialias=False): half-pixel centers, no antialias."""
    h, w = arr.shape[:2]

    def axis(n_out, n_in):
        src = np.clip((np.arange(n_out) + 0.5) * (n_in / n_out) - 0.5, 0, n_in - 1)
        lo = np.floor(src).astype(int)
        return lo, np.minimum(lo + 1, n_in - 1), (src - lo).astype("float32")

    y0, y1, wy = axis(size, h)
    x0, x1, wx = axis(size, w)
    top = arr[y0][:, x0] + (arr[y0][:, x1] - arr[y0][:, x0]) * wx[None, :, None]
    bot = arr[y1][:, x0] + (arr[y1][:, x1] - arr[y1][:, x0]) * wx[None, :, None]
    return top + (bot - top) * wy[:, None, None]

sess = ort.InferenceSession("backbone.onnx")
H = dict(np.load("gradcam_head.npz"))

img = Image.open("xray.jpg").convert("RGB")
arr = resize_bilinear(np.asarray(img, dtype="float32"))          # NOT img.resize(...)
x = (arr / 255.0 - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225]
conv = sess.run(None, {sess.get_inputs()[0].name: x[None].astype("float32")})[0][0]

gap = conv.mean((0, 1))
a1 = np.maximum(gap @ H["fc1_W"] + H["fc1_b"], 0)
n1 = H["bn1_gamma"] * (a1 - H["bn1_mean"]) / np.sqrt(H["bn1_var"] + H["bn1_eps"]) + H["bn1_beta"]
a2 = np.maximum(n1 @ H["fc2_W"] + H["fc2_b"], 0)
n2 = H["bn2_gamma"] * (a2 - H["bn2_mean"]) / np.sqrt(H["bn2_var"] + H["bn2_eps"]) + H["bn2_beta"]
prob = 1 / (1 + np.exp(-(n2 @ H["prob_W"][:, 0] + H["prob_b"][0])))
print("P(pneumonia) =", float(prob))

Raw sigmoid output is uncalibrated. For the full pipeline as deployed, apply temperature scaling (temperature.json, T = 0.72) and then the calibrated operating threshold and abstention band from triage.json — see api/inference.py.