resnet50-chest-xray-256-fullframe-gray

ResNet50: chest X-ray pneumonia (256x256 full-frame grayscale)

Binary classifier (NORMAL vs PNEUMONIA): torchvision ResNet50 (ImageNet IMAGENET1K_V2 weights) fully fine-tuned on Purino/chest-xray-pneumonia-256-fullframe-gray. Research/educational use only, not a diagnostic tool.

Evaluation (official test split, 624 images, evaluated once with the best-validation checkpoint)

Per class (threshold 0.5):

Confusion matrix (rows = true, columns = predicted):

Validation (15% stratified hold-out of train, 783 images) at the selected epoch 8: accuracy 0.9885, recall 0.9880, macro-F1 0.9851, AUC 0.9983.

Training

• Data: official train split (5216 images) -> 4433 train / 783 val (stratified, seed 42); official val split unused (too small); test held out.
• Input: 256x256 grayscale replicated to 3 channels, ImageNet mean/std. Augmentation: random affine (±10°, 5% shift, 0.9-1.1 scale), brightness/contrast jitter, no flips.
• Head: Dropout(0.3) + Linear(2048, 2). Loss: class-weighted cross-entropy (train class counts {'NORMAL': 1140, 'PNEUMONIA': 3293}).
• Optimizer: AdamW (lr 0.0001, weight decay 0.0001), cosine schedule, batch 64, up to 20 epochs, mixed precision, 2 GPU(s).
• Checkpoint: best validation macro-F1 (epoch 8), early-stopping patience 5.

Usage

import torch, torch.nn as nn
from torchvision import models, transforms as T
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
from PIL import Image

m = models.resnet50(); m.fc = nn.Sequential(nn.Dropout(0.3), nn.Linear(2048, 2))
m.load_state_dict(load_file(hf_hub_download("Purino/resnet50-chest-xray-256-fullframe-gray", "model.safetensors"))); m.eval()
tf = T.Compose([T.Resize((256, 256)), T.ToTensor(), T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
x = tf(Image.open("xray.png").convert("RGB")).unsqueeze(0)
with torch.no_grad(): p = m(x).softmax(1)[0]
print(["NORMAL", "PNEUMONIA"][p.argmax().item()], p.tolist())

Limitations

• The dataset has no patient-level split or deduplication, so validation (and possibly test) images may come from patients seen in training; scores can be optimistic.
• Trained on one pediatric dataset (Kermany et al.); it will likely not generalize to other scanners, populations or adult chest X-rays without further validation.
• Images were resized to 256x256 without cropping, which changes aspect ratio. Labels were not independently verified here.

Data & credits

Source: Kermany et al. (Mendeley, CC BY 4.0) via Paul Mooney's Kaggle upload, converted to full-frame grayscale 256x256.