AI-Powered Radiomics Algorithm Based on Slice Pooling for the Glioma Grading

Slice Pooling AI Model

AI-Assisted Radiomics Algorithm for Glioma Grading Based on Slice Pooling

Background Introduction

Glioma is the most common and threatening tumor in the central nervous system, characterized by high incidence, high recurrence rates, high mortality, and low cure rates. The World Health Organization (WHO) classifies gliomas into four grades (I, II, III, and IV), where grades I and II are called low-grade gliomas (LGG), and grades III and IV are referred to as high-grade gliomas (HGG). High-grade gliomas are more aggressive malignant tumors with an average life expectancy of about two years. Although the WHO introduced molecular typing in 2016 to exclude insensitive treatments, grading remains an important diagnostic criterion for gliomas as it determines treatment options.

Magnetic Resonance Imaging (MRI) is a commonly used imaging technique for detecting and analyzing gliomas. It is a non-invasive and rapid method, and MRI images contain rich information that is difficult to obtain through mere observation by doctors. Radiomics, as part of artificial intelligence technology, has become a popular quantitative imaging analysis method. It can quantify the extracted image information into features that assist in diagnosing and prognosticating various diseases. The preliminary results in clinical research in this field have made people highly anticipate the transition of radiomics into clinical practice through high-performance computer-aided diagnosis (CAD) systems.

CAD systems provide integrated solutions, especially for image-based solutions, not limited to medical and biological problems. They enable accurate preoperative diagnosis based on imaging examinations. For example, radiologists need only segment the Region of Interest (ROIs) on MRI images, then input them into the CAD system, which will generate diagnostic results, including pathological details. This new diagnostic model provides analysis results at least on par with expert diagnoses and enhances diagnosis speed. In summary, it offers both reference and time-saving for choosing patient treatment plans.

Despite the promising performance of radiomics-based CAD systems, there is still room for improvement. Currently, ROI segmentation is performed semi-automatically or manually. In MRI images of tumor diseases, experienced radiologists manually segment the ROIs in each slice, leading to substantial labor and time consumption, delaying the entire CAD system workflow. Therefore, an appropriate solution is needed to alleviate the radiologists’ burden of manual ROI delineation.

Paper Source

This paper, authored by Guohua Zhao (First Affiliated Hospital of Zhengzhou University, Collaborative Innovation Center of Internet Healthcare of Zhengzhou University), Panpan Man, Jie Bai (First Affiliated Hospital of Zhengzhou University), Longfei Li, Peipei Wang (First Affiliated Hospital of Zhengzhou University), Guan Yang (School of Computer Science, Zhongyuan University of Technology), Lei Shi (School of Software, Zhengzhou University), Yongcai Tao, Yusong Lin (School of Software, Zhengzhou University, Collaborative Innovation Center of Internet Healthcare of Zhengzhou University, and HANWEI IOT Research Institute), and Jingliang Cheng (First Affiliated Hospital of Zhengzhou University), was published in the IEEE Transactions on Industrial Informatics in August 2022.


This study established an AI-powered radiomics algorithm based on slice pooling, named AI-RASP, to improve the efficiency and accuracy of glioma grading.

Methods and Process

1. Slice Pooling (SP) Mechanism

The Slice Pooling (SP) mechanism is an essential preprocessing step of AI-RASP. Similar to the pooling mechanism in deep learning models, SP is a pixel-based compression mechanism. This mechanism merges pixel values at the same position in different slices, choosing the maximum value as the pixel value for that position to compress the image. The specific formula is as follows:

[ I_s (i, j) = \max(S_1(i, j), \ldots, S_k(i, j), \ldots, S_l(i, j)) ]

Here, ( S_k (i, j) ) represents the pixel value at position (i, j) in the k-th slice, and ( I_s (i, j) ) represents the pixel value at position (i, j) in the compressed image.

2. AI-RASP Workflow

The AI-RASP algorithm we proposed is a prototype of a CAD system that can be used for glioma grading. AI-RASP includes the following steps:

(i) Image Preprocessing

First, the images of four MRI sequences (T1-weighted imaging, T2-weighted imaging, enhanced T1-weighted imaging, and FLAIR imaging) are input into the radiomics model. Each sequence undergoes the SP operation to generate compressed images. All images are preprocessed before being input into the model, including skull stripping using the FSL tool, voxel normalization, and gray-scale normalization.

(ii) Segmentation and Registration

Before segmentation and registration, three preprocessing groups are set up for comparison. Group 1 consists of original images manually segmented by radiologists; Group 2 comprises compressed images generated by SP and manually segmented compressed ROIs; Group 3 consists of compressed images generated by SP, manually segmented by radiologists. The 2-D ROIs of all tumors are manually segmented by two experienced radiologists using ITK-SNAP software and reviewed by a senior radiologist with over 20 years of experience to ensure segmentation accuracy.

(iii) Radiomics Feature Extraction and Selection

Using the pyradiomics software package of Python, radiomics features are extracted and selected from the original and compressed images. The feature selection methods include the Mann-Whitney U test, elastic net method, and recursive feature elimination algorithm to avoid overfitting of current features.

(iv) Model Construction and Validation

A linear Support Vector Machine (SVM) model predicts glioma grades based on features selected in the training set, and the model is validated in different preprocessing groups. ROC curve analysis includes calculating AUC, sensitivity, and specificity values to evaluate model accuracy.


Results and Analysis

The proposed algorithm was validated on a multi-center dataset involving 400 patients, showing significant improvement in segmentation efficiency (more than five times faster) compared to traditional manual segmentation and consistent performance in glioma grading accuracy. Specifically, the AUC values in the training set reached 0.927, and those in the validation set reached 0.896, 0.902, and 0.894, respectively.

Significance and Value

The application of AI-RASP in glioma grading significantly reduces the time radiologists need for manual ROI segmentation, enhancing work efficiency. The feature extraction from compressed images performs well, aiding the clinical application of radiomics. Considering its significant value in scientific research and its broad application prospects in enhancing clinical diagnostic efficiency and accuracy, AI-RASP is highly valuable.

Conclusion

This paper proposes an AI-driven radiomics algorithm based on slice pooling (AI-RASP), significantly improving segmentation efficiency and accuracy in glioma imaging grading. Future research will further validate the effectiveness of AI-RASP in advanced MRI techniques, more diverse histological examinations, and prospective data.