# Artificial Intelligence and Rectal Cancer: Beyond Images

## TL;DR

This paper argues that artificial intelligence in rectal cancer care should go beyond images, using data like health records and omics for better results.

## Contribution

The paper is the first to review non-image AI models in rectal cancer, highlighting their potential and calling for more research.

## Key findings

- Non-image AI models perform well and are under-researched compared to image-based models.
- Combined models using multiple data types often outperform single-modality models.
- Multicenter studies on non-image and combined models are scarce but needed for validation.

## Abstract

The cancer burden, particularly in rectal cases, can be alleviated through the use of artificial intelligence models, provided they are properly designed, implemented, and validated. Artificial intelligence encompasses machine learning, which in turn includes deep learning. Artificially intelligent models can be developed based on various types of data, including images, numerical values, and texts. We believe there is currently considerable hype around image-based models, and that more intensive exploration of other data types—such as electronic health records and omics—could greatly enhance both research and clinical practice. By analyzing the literature, we confirm this idea and offer some recommendations that we ultimately consider beneficial for patients, especially by promoting multimodal approaches beyond simply imaging.

Introduction: The variability of cancers and medical big data can be addressed using artificial intelligence techniques. Artificial intelligence models can accept different input types, including images as well as other formats such as numerical data, predefined categories, and free text. Non-image sources are as important as images in clinical practice and the literature; nevertheless, the secondary literature tends to focus exclusively on image-based inputs. This article reviews such models, using non-image components as a use case in the context of rectal cancer. Methods: A literature search was conducted using PubMed and Scopus, without temporal limits and in English; for the secondary literature, appropriate filters were employed. Results and Discussion: We classified artificial intelligence models into three categories: image (image-based input), non-image (non-image input), and combined (hybrid input) models. Non-image models performed significantly well, supporting our hypothesis that disproportionate attention has been given to image-based models. Combined models frequently outperform their unimodal counterparts, in agreement with the literature. However, multicenter and externally validated studies assessing both non-image and combined models remain under-represented. Conclusions: To the best of our knowledge, no previous reviews have focused on non-image inputs, either alone or in combination with images. Non-image components require substantial attention in both research and clinical practice. The importance of multimodality—extending beyond images—is particularly relevant in the context of rectal cancer and potentially other pathologies.