DECIMER Image Classifier

• This model aims to classify whether or not an image is a chemical structure. It was built using EfficientNetB0 as a base model, using transfer learning and fine-tuning it.

• It gives a prediction between 0 and 1 where 0 means it is a chemical structure and 1 means it is not. The data used to build the model is available in Zenodo.

Performance

The model was trained on 10905114 images, validated on 2179798 images and tested on 544946 images. It took 52h and 15 minutes and 17GB on a Tesla V100-PCI-E-32GB GPU.

From the results on the test set, we computed the AUC and the Youden index (on the image below). Using the prediction threshold marked by the Youden index, 0.000089, we achieved the following performance metrics:

MCC = 0.989944

Accuracy = 0.994968

Sensitivity = 0.993049

Specificity = 0.996888

Model construction

The model construction script is available on the model_construction folder with comments for every step taken.

Installation

• Simply run pip install git+https://github.com/Iagea/DECIMER-Image-Classifier

Usage

from decimer_image_classifier import DecimerImageClassifier

# Instantiate DecimerImageClassifier (this loads the model and all needed settings)
decimer_classifier = DecimerImageClassifier()

# If you just need a statement about the image at image_path being a chemical structure or not:
result = decimer_classifier.is_chemical_structure(image_path)

# If you want the classification score (value between 0 (-> chemical structure) and 1 (-> no chemical structure))
score = decimer_classifier.get_classifier_score(image_path)

Example

An example with 10 images from the test set that will be available in Zenodo is also present in the form of a Jupyter notebook in the folder examples.

Author: M. Isabel Agea

Project Website:

• A web application implementation is available at decimer.ai, implemented by Otto Brinkhaus

Research Group