Publisher Correction: Converting tabular data into images for deep learning with convolutional neural networks

Yitan Zhu, Thomas Brettin, Fangfang Xia, Alexander Partin, Maulik Shukla, Hyunseung Yoo, Yvonne A. Evrard, James H. Doroshow, Rick StevensScientific 2021 · DOI

This document serves as a formal correction to the research paper "Converting tabular data into images for deep learning with convolutional neural networks," originally published in Scientific Reports in May 2021. The correction addresses a specific typographical error in the indexing of the Image Generator for Tabular Data (IGTD) algorithm.

Context and Algorithm Overview

The original research introduced the IGTD (Image Generator for Tabular Data) algorithm, a novel method designed to bridge the gap between non-spatial tabular data and deep learning architectures like Convolutional Neural Networks (CNNs). While CNNs are highly effective at identifying spatial correlations in images, tabular data typically lacks this inherent structure.

The IGTD algorithm solves this by mapping features from a tabular dataset onto a 2D pixel grid. By arranging features such that similar attributes are placed near one another, the algorithm creates a spatial structure that allows CNNs to leverage their filtering and feature-extraction capabilities on data that was previously unstructured.

Description of the Correction

This correction addresses a clerical error in the formal presentation of the algorithm's pseudocode. Specifically, the indexing of the "Otherwise" branch in the algorithm description was incorrectly sequenced in the original manuscript, which could potentially lead to confusion during technical implementation or replication.

The update corrects the following labels:

  • Original (Incorrect): Steps were labeled (xxii), (xxiii), and (xxiv).
  • Corrected: Steps are now labeled (v), (vi), and (vii).

These steps represent the iterative assignment and state-tracking logic necessary for the feature-to-pixel mapping process, specifically: $h_{n^*} = s$, $e_s = e_{s-1}$, and $k_s = k_{s-1}$. This change aligns the steps with the proper logical flow of the algorithm.

Significance for the Research Community

Accuracy in algorithmic descriptions is paramount for scientific reproducibility. Even small typographical errors in indexing can hinder the ability of engineers and researchers to correctly implement complex logic. By updating the scientific record, the authors ensure that the mathematical description of the IGTD algorithm is consistent, accurate, and ready for use in future deep learning applications.

The original article has been updated to reflect these changes. The core methodology, results, and conclusions of the 2021 study remain unaffected by this correction.

Summary updated

BibTeX
@inproceedings{zhu2021publisher,
  author = {Yitan Zhu and Thomas Brettin and Fangfang Xia and Alexander Partin and Maulik Shukla and Hyunseung Yoo and Yvonne A. Evrard and James H. Doroshow and Rick Stevens},
  title = {{Publisher Correction: Converting tabular data into images for deep learning with convolutional neural networks}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-93376-5},
}
Figure