Author: Weigang Hu, Zhenhao Li, Jiazhou Wang, Xiaojie Yin, Zhen Zhang 👨🔬
Affiliation: Fudan University Shanghai Cancer Center 🌍
Purpose:
This study aims to develop and validate a novel deep learning method to generate synthetic PET images for rectal cancer from MRI data. By incorporating metabolic information from the synthetic PET images, we seek to improve tumor assessment, offering new approaches for accurate diagnosis and prognosis prediction in rectal cancer.
Methods:
A deep learning model was developed to synthesize PET images from MRI data. The model was trained on data from 150 patients with locally advanced rectal cancer, using baseline MR and PET/CT images acquired within one month. Model performance was evaluated by comparing features of synthetic and real PET images from 21 patients, focusing on key metrics such as metabolic tumor volume at SUV thresholds of 4 (MTV4) and 6 (MTV6), as well as SUVmax and SUVmean. Two additional datasets, comprising 392 and 346 patients, were used to assess the clinical value of the synthetic PET images in overall survival (OS) prediction and pathological complete response (pCR) evaluation. Additionally, an external validation was conducted using PET/MR images from three rectal cancer patients from another institution.
Results:
Correlation coefficients between synthetic and real PET images for MTV4 and MTV6 were 0.82 and 0.75. In terms of OS prediction, indicators such as MTV4 and MTV6 emerged as independent prognostic factors, with a hazard ratio (HR) of 1.76 (95% CI: 1.24-2.51) for MTV4 (p=0.004) and 1.56 (95% CI: 1.09-2.21) for MTV6 (p=0.032). The AUC for predicting pCR based on clinical indicators was 0.72, which increased to 0.84 when combined with synthetic PET features. External validation showed SSIM and PSNR of MR-synthetic PET against concurrently acquired PET images were 0.79 and 24.59 dB.
Conclusion:
Synthesizing PET images from MRI generated by deep learning model shows improvement in OS prediction and pCR assessment to support personalized treatment decisions.