pcam-densenet121-cnn

PCam DenseNet121 CNN (Fold 0)

This repository contains a TensorFlow/Keras DenseNet121-based CNN trained on the PatchCamelyon (PCam) dataset for binary histopathology image classification (tumor vs. normal).

This model corresponds to fold 0, which performed best in a 3-fold ensemble.

Model details

• Framework: TensorFlow / Keras (nightly)
• Base model: tf.keras.applications.DenseNet121
• Input shape: 96 × 96 × 3 RGB
• Output: Single sigmoid unit (probability of tumor)
• Loss: Binary cross-entropy

The file pcam_advanced_ensemble_model0_best.keras is a full Keras model (architecture + weights).

How to use

from huggingface_hub import hf_hub_download
import tensorflow as tf
from tensorflow.keras.applications.densenet import preprocess_input
import numpy as np

# Download model file from the Hub
model_path = hf_hub_download(
    repo_id="NSH0407/pcam-densenet121-cnn",
    filename="pcam_advanced_ensemble_model0_best.keras"
)

# Load model (note custom_objects for preprocess_input)
model = tf.keras.models.load_model(
    model_path,
    custom_objects={"preprocess_input": preprocess_input}
)

# img: np.array of shape (96, 96, 3), values in [0, 255] or [0.0, 255.0]
img = ...  # TODO: load your own image here

x = np.expand_dims(img.astype("float32"), axis=0)
x = preprocess_input(x)

prob = float(model.predict(x)[0, 0])
print("Tumor probability:", prob)