Deep Learning Application in Diagnosing Breast Cancer Recurrence Publisher



Jam Z¹ ; Albadvi A¹ ; Atashi A^{2, 3} Show Affiliations
Source: Multimedia Tools and Applications Published:2025

Abstract

Patients' lives can always be saved when diseases, especially special diseases, are detected early. The chances of a patient surviving can be increased by early detection. Breast cancer is one of the deadliest and common cancers. After recovering from breast cancer, patients are always worried about recurrence and return. The use of modern technology, however, can help predict disease recurrence at an early stage, allowing patients to receive treatment sooner. Significant strides have been achieved in deep learning, demonstrating strong performance in handling unstructured data challenges. However, when it comes to predicting tabular data, deep learning hasn't quite matched its success with unstructured data. Presently, ensemble models relying on gradient-boosted decision trees (GBDT) are frequently favored for tabular data prediction tasks. Typically, these GBDT-based models outshine deep learning approaches. Many novel deep learning techniques are emerging for handling tabular data. TabNet, for instance, mirrors decision tree feature selection within a neural network framework. AutoInt addresses high dimensionality by condensing data through embedding layers. Tab Transformer adapts the transformer model, generating text representations for categorical attributes. Despite their innovation, these methods remain less recognized compared to those for image and text data processing. In this study, 158 different characteristics of 5142 breast cancer patients from 1997 to 2019 were examined. We aim to evaluate deep learning techniques effectiveness in detecting breast cancer recurrence. Through examination of evaluation metrics, it becomes evident that deep learning approaches applied to tabular data surpass traditional machine learning algorithms, even when dealing with imbalanced datasets. Ultimately, the results derived from each algorithm analyzed and concluded with a review and comparison of the findings. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.