Comprehensive study comparing different machine learning methods in computed tomography imaging

Abstract

The article written by Zhao et al, which was recently accepted for publication, introduces an innovative method that combines deep learning-based feature extraction with a radiomics nomogram to create a noninvasive procedure for determining perineural invasion status in patients with rectal cancer. This method is an artificial intelligence application in which researchers segment their own datasets, derive features and analyze their weights. It was found that the support vector machine was the most effective model in the arterial and venous phases. A support vector machine is a machine learning algorithm based on a vector space that finds a decision boundary between the two classes furthest from any point in the training data.

Key Words: Deep learning; Perineural invasion; Radiomics; Rectal cancer; Stacking nomogram; Support vector machines

Core Tip: This review is about the article written by Zhao et al. This study compares different machine learning methods in computed tomography imaging. In this study support vector machines, a vector space-based machine learning algorithm that finds a decision boundary between the two classes that are farthest from any point in the training data, was found to be the most effective model in the arterial and venous phases.

INTRODUCTION

The article written by Zhao et al[1], which was recently accepted for publication, introduces an innovative method that combines deep learning-based feature extraction with a radiomics nomogram to create a noninvasive procedure for determining perineural invasion (PNI) status in patients with rectal cancer. This method is an artificial intelligence (AI) application in which researchers segment their datasets, derive features, and analyze their weights. As a result, support vector machines, a vector space-based machine learning algorithm that finds a decision boundary between the two classes that are farthest from any point in the training data, were the most effective model in the arterial and venous phases. In other studies using anonymous datasets, different machine-learning algorithms were more effective[2,3]. PNI presence significantly affects the prognosis of individuals with rectal cancer in the long term[4]. It is important to evaluate the benefit of postoperative chemotherapy and neoadjuvant chemo-radiation treatment of patients. The heterogeneity of invasion reduces the accuracy of invasive diagnostic methods and leads to complications. For this reason, it is very important to determine the PNI status more precisely by a non-invasive method before the operation. Although deep learning is a well-established technique, its application to certain niche areas, as shown in this study, can provide new insights.

APPLICATION OF DEEP LEARNING FEATURE EXTRACTION BASED ON PRE - TRAINED IMAGENET MODEL IN MEDICAL IMAGING AND FUTURE RESEARCH DIRECTIONS

The main contribution of the article to the literature is that the researchers used transfer learning with a pre-trained ImageNet model for feature extraction. Deep convolutional neural networks were used based on pre-trained weights on ImageNet to extract deep learning features. The article addresses how the data are adapted for the pre-trained model, such as matching the pre-processing used for ImageNet images, applying similar normalization techniques, and using all layers for feature extraction. ResNet-50 was selected as the backbone network of the deep convolutional neural network model due to its pre-training features on ImageNet, and it is stated that it provides rich convolutional layer features that have a good ability to represent complex tumor regions as a region of interest in computed tomography images. The article evaluates the performance of the model based on the area under the curve metric, while also providing details on other metrics such as precision, recall, or F1 score.

Future studies on this topic can include data from different imaging modalities, such as magnetic resonance imaging and/or positron emission tomography scans, to enhance predictive performance by using complementary information. Additionally, the study can benefit from external validation across diverse patient populations to assess the generalizability and robustness of the model in other clinical settings. Another important area for future development is the interpretability of deep learning models in radiomics. Studies primarily focus on predictive performance metrics, but incorporating explainability techniques such as saliency maps or Shapley Additive Explanations can help radiologists and clinicians better understand how specific imaging features contribute to the model’s decision-making process[5]. Increasing transparency will increase confidence in AI-driven diagnostics and facilitate its adoption in clinical practice.

CONCLUSION

I think this very comprehensive original study will lead to new studies in different applications of AI and machine learning in the field of diagnostic radiology.