Towards More Accurate Chemical Image Recognition: Bridging ML, OSRA, and the Chemical Image Classifier

Aleksei Krasnov

The field of Optical Chemical Structure Recognition (OCSR) has recently seen renewed interest due to advancements in Deep Learning and the development of specialized image recognition models. This study investigates the effectiveness of various OCSR methods, including DECIMER, MolScribe, RxnScribe, and OSRA, using an independent dataset derived from patents and patent applications.

Our comparative analysis reveals that MolScribe and RxnScribe outperform OSRA in extracting single structure and reaction information from images, particularly those involving small organic molecules. While DECIMER and MolScribe yield similar but superior results for single small molecule structures compared to OSRA, OSRA demonstrates its strengths in handling images of large molecules, transition metal complexes, and those containing multiple compound structures.

To optimize the process of chemical image recognition, we developed the Chemical Image Classifier (ChemIC), a machine learning-based tool designed to categorize images by chemical modality and route them to the most appropriate OCSR method. This approach leverages the strengths of each method according to the specific characteristics of the images, significantly improving the accuracy and reliability of chemical structure extraction.

Despite the advances achieved with current AI-based OCSR methods, the study identifies several areas for future development, such as enhanced image-resizing techniques for larger molecules, and the need for more comprehensive training datasets that include complex molecules like oligomers, polymers, and metal-organic structures.

Additionally, we advocate for future OCSR methods to support V2000 or V3000 files as output formats, facilitating the accurate prediction of more complex Markush and polymeric structures in a universally accepted chemical structure format. The continuous refinement of OCSR methods, guided by these insights, will be crucial in addressing the current limitations and pushing the boundaries of chemical image recognition.