1. Introduction

CEST is a promising and powerful tool that allows sensitively detection of metabolic changes caused by diseases such as cancer and ischemia¹. Current clinical CEST protocols, termed as Amide proton transfer(APT), could only derive one contrast map (APTw) using asymmetric analysis. More dedicate analysis could be achieved by fitting the lineshape of Z-spectra, but require densely-sampled saturation offsets. Previously we introduced a model-free analysis method, featuring fully-usage of acquired saturation frequency by treating the acquired CEST series from each voxel as a sample, and then ranking the frequency importance according to a permute random forest (PRF) algorithm². This method proved useful in predict H3K27 gene expression in glioma patients³. Therefore, there is a need for an inline software on scanner, which could be used friendly right after the scans and may facilitate the CEST applications.

2. Methods

2.1 Permuted random forest (PRF) on CEST
Permuted random forest is a machine learning-based method, which trains and predicts samples using multiple trees. In PRF-CEST, we label lesion group as “1” and control group as “0” by drawing ROIs manually or loading a ROI file, then the sampled Z-spectra data is used for permuted random forest analysis.
2.2 Design and Implementation of the Software
This software is compiled using Python 3.9 with PyQt5 and runs in PRIDE 2.0 on Philips scanner, embedded in the construction we developed for post-processing and visualization of CEST (PV-CEST)⁴. A page was designed to display medical images and draw ROIs. The main page of PRF-CEST and the workflow is shown in Fig.1&2. As seen in Fig.2, the workflow includes image viewing, ROI drawing, and Data analysis. Researchers can load and observe images on this page, draw ROIs on lesion group and control group respectively, and perform permuted random forest on data. Notably, PRF-CEST can be compared and analyzed either on a single image or multiple images. After the analysis is completed, the frequency feature importance ranking chart is displayed and saved.
Fig. 3&4 displays the verification of this software. Fig. 3 shows the case of comparison on a single image, that is between disease areas and normal areas in the same brain layer of recurrent glioblastoma patient, with maximum contribution at 3.5ppm. Fig. 4 shows the case of comparison on multiple images, a newborn rat’s brain layer at 1th week and at 5th week, with maximum contribution at 3.5ppm. This is consistent with the result of LDamide maps, which indicates PRF could become a powerful clinical tool.
2.3 Workflow
2.3.1 Image viewing
After entering the software homepage, select the frequency file and open the CEST data file, then enter the PRF-CEST interface to display the image and load the data.
2.3.2 ROI drawing
Draw the ROI of the lesion area/normal area manually on the image. When the left mouse button is pressed, the drawing of ROI automatically tracks the movement of the mouse. After releasing the left mouse button, the start and end points of ROI are automatically connected, and points inside the ROI are selected for recording. In the process, users are allowed to use right mouse button to clear the drawn ROI, import/export ROI or clear all the data recorded. The Z-spectrum data of the ROI area after drawing is recorded.
2.3.3 Data analysis
Perform PRF analysis on the data and an image would be displayed after performing the permuted random forest algorithm excluding frequency offset between -0.5ppm and 0.5ppm. Users are able to choose the frequency offset needed to display and perform comparative analysis on single or multiple images by loading various images.

3. Discussion

We have developed an inline model-free software tool that can directly use permuted random forest to analyze CEST frequency importance and validated the functionality of the software using rat brain data. Experimenters could draw ROIs to choose areas as needed and perform comparative analysis either on single images or on multiple images from various subjects by using raw data without fitting, which improves the efficiency of data analysis. The way of manually drawing ROI improves the flexibility and operability of data. In the future, we can further cooperate with hospitals and other institutions to play a role in practical fields.

4. Conclusion

PRF-CEST provides a method of drawing ROI manually to perform permuted random forest analysis for frequency importance, which accelerates analysis efficiency, facilitates researchers to make judgments, and has potential applications.

Acknowledgements

This work is partially supported by National Key R&D Program of China 2022YFC3602500，2022YFC3602503 and National Natural Science Foundation of China (NSFC) (Nos. 82071914).