PCNSL

This is code of paper: Deep Learning-based Overall Survival (OS) Prediction Model in Patients with Rare Cancer: A Case Study for Primary Central Nervous System Lymphoma (PCNSL), published on International Journal of Computer Assisted Radiology and Surgery (2023).

Link of paper: She, Ziyu, et al. "Deep learning-based overall survival prediction model in patients with rare cancer: a case study for primary central nervous system lymphoma." International Journal of Computer Assisted Radiology and Surgery (2023): 1-8. https://link.springer.com/article/10.1007/s11548-023-02886-2

Keypoints of paper:

1. Data Collection and Data Preprocessing

   

   Data Collection: T1-weighted (T1), T2-weighted (T2), post-contrast T1-weighted (T1Gd) and OS of these patients are collected. OS positive means OS more than 1 year and negative means less than 1 year. There are 30 positive cases and 26 negative cases.

   Step #1: Bias field correction, registration, skull stripping, intensity normalization and voxel resampling using 3D Slicer.

   Step #2: Data augmentation including background removal, elastic deformation, random spatial cropping, random rotation and resizing using MONAI (preprocessing_resnet.py).

2. 3D Residual Network (ResNet), Transfer Learning and Gradient-weighted Class Activation Mapping (Grad-Cam)

   

   3D ResNet Architecture: Input 3D images, use convolution layers and pooling layers as encoder to extract features, finally, use a linear FC layer to output OS classification results (train_resnet.py).

   Transfer Learning: Pre-train 3D ResNet on BraTS2020 for OS classification of patients with glioblastoma (BraTS2020 consists of multi-modal preoperative images of 235 glioblastoma patients from 19 institutions with OS), then, fine-tune the 3D ResNet on PCNSL dataset for the target task OS classification of patients with PCNSL.

   Grad-cam: Output pattern of the 3D ResNet (visualization_resnet.py).

3. Result

   

   Qutitative Result: Cross-validation results of ML models and 3D ResNet. SVM (Support Vector Machine). T (Training from Scratch). TL (Transfer Learning). Clinic means clinical data. T1r, T2r and T1Gdr mean radiomics data from T1, T2 and T1Gd. * means p < 0.05 in the t-test. Best performance on T1Gd, consistent with clinical outcome.

   

   

   Qulitative Result: OS pattern of 3D ResNet. #number (patient ID). MRI (TIGd slice near tumor). Activation Map (activation map of the slice). Colorbar shows activation map intensity. PCNSL is a whole-brain disease; in cases where OS less than 1 year (Bottom), it is more difficult to distinguish tumor boundary from normal part of the brain than where OS more than 1 year (Top); consistent with clinical outcome.