ConvNeXt Lung Cancer Histopathology Classifier

This is a model card intended to be used as the README.md of the Hugging Face model repository:

https://huggingface.co/felixhrdyn/convnextv1-lung-cancer

The model is a fine-tuned ConvNeXt-Base classifier for lung histopathology tissue type prediction.

Model Details

Model type: Image classifier (3 classes)
Base architecture: ConvNeXt-Base (pretrained on ImageNet-1K)
Framework: TensorFlow / Keras
Model file: convnext_lung_82.keras
Input: RGB image resized to 224×224, scaled to [0, 1]
Output: Softmax probabilities for 3 tissue classes

Labels

The model predicts one of the following:

Adenocarcinoma
Benign Tissue
Squamous Cell Carcinoma

Intended Use

Educational and research use for lung histopathology image classification.
Prototyping and benchmarking on data that is similar to the training distribution.

Out-of-Scope Use

Clinical diagnosis or patient-care decisions.
Use on data distributions significantly different from the training dataset without validation.

Training Data

Dataset: LC25000 Lung and Colon Histopathology Dataset (lung subset)
Subset used: 3 lung classes with 1,000 images per class (3,000 total)

This model card summarizes the training/evaluation reported by the LUCIAN project.

Training Procedure

Transfer learning using ConvNeXt-Base pretrained weights.
A small classification head is added and the model is fine-tuned.
The “best checkpoint” is selected by maximum val_accuracy.

Evaluation

Two split strategies were evaluated in the LUCIAN project:

Metric	Split 80:10:10 (Final)	Split 70:15:15
Train Accuracy	96.08%	94.95%
Validation Accuracy	96.67%	94.00%
Test Accuracy	93.67%	90.44%
Precision (macro)	93.63%	90.47%
Recall (macro)	93.67%	90.44%
F1-Score (macro)	93.64%	90.39%

Metrics are reported from the best checkpoint epoch selected by ModelCheckpoint(monitor='val_accuracy', mode='max').

How to Use

Load model from the Hub

from huggingface_hub import hf_hub_download
import tensorflow as tf

path = hf_hub_download(
    repo_id="felixhrdyn/convnextv1-lung-cancer",
    filename="convnext_lung_82.keras",
)
model = tf.keras.models.load_model(path)

Preprocess and run inference

This matches the preprocessing used in the LUCIAN app (resize(224, 224), RGB, normalize to [0, 1]).

import numpy as np
from PIL import Image
import tensorflow as tf

IMAGE_SIZE = (224, 224)
LABELS = ["Adenocarcinoma", "Benign Tissue", "Squamous Cell Carcinoma"]

img = Image.open("your_image.jpg").convert("RGB").resize(IMAGE_SIZE)
x = np.expand_dims(np.array(img) / 255.0, axis=0).astype(np.float32)

probs = model(tf.convert_to_tensor(x), training=False).numpy()[0]
pred = LABELS[int(np.argmax(probs))]
print(pred, probs)

Limitations and Risks

The model was trained on a specific dataset and preprocessing pipeline; performance may degrade on different scanners, staining protocols, magnifications, or patient populations.
Like most CNNs, it can be sensitive to artifacts (blur, compression, pen marks) and distribution shift.
This is not a medical device and has not undergone clinical validation.

Ethical Considerations

Do not use as a substitute for expert pathology assessment.
If used in downstream workflows, ensure appropriate human oversight, dataset audits, and validation.

Citation

If you use this model, cite the LUCIAN project (and the ConvNeXt paper):

Liu et al., “A ConvNet for the 2020s” (ConvNeXt)

Downloads last month: 27

felixhrdyn
/

convnextv1-lung-cancer