Source-linked AI summary
Prostate Cancer Diagnosis using Deep Learning with 3D Multiparametric MRI
Saifeng Liu, Huaixiu Zheng, Yesu Feng, Wei Li
TL;DR
Prostate cancer diagnosis is limited by imperfect imaging assessment and variable engineered-feature models. This paper develops and end-to-end trains XmasNet on 3D multiparametric MRI using augmentation that incorporates lesion views. XmasNet outperformed engineered-feature machine-learning models and achieved the second-highest AUC in the PROSTATEx challenge.
Problem
Existing prostate cancer screening and mpMRI assessment have limited specificity or variable accuracy, while engineered-feature models depend substantially on domain expertise and feature quality.
Method
XmasNet is an end-to-end convolutional neural network trained on 3D multiparametric MRI with 3D rotation and slicing-based augmentation.
Results
Second-highest AUC in the 2017 PROSTATEx challenge was achieved, while XmasNet outperformed engineered-feature conventional machine learning on both train and test data.
Takeaways & Limitations
The study demonstrates the potential of deep learning for cancer imaging and prostate cancer diagnosis.
Abstract
from arXiv · showhide
A novel deep learning architecture (XmasNet) based on convolutional neural networks was developed for the classification of prostate cancer lesions, using the 3D multiparametric MRI data provided by the PROSTATEx challenge. End-to-end training was performed for XmasNet, with data augmentation done through 3D rotation and slicing, in order to incorporate the 3D information of the lesion. XmasNet outperformed traditional machine learning models based on engineered features, for both train and test data. For the test data, XmasNet outperformed 69 methods from 33 participating groups and achieved the second highest AUC (0.84) in the PROSTATEx challenge. This study shows the great potential of deep learning for cancer imaging.
1. INTRODUCTION
Prostate cancer diagnosis remains challenging because existing screening and mpMRI-based assessment have limited or variable accuracy. The paper motivates deep learning as an alternative to expertise-dependent engineered features for 3D mpMRI diagnosis.
- PSA and DRE screening have limited diagnostic accuracy and may cause overdiagnosis.
- mpMRI with PI-RADS improves sensitivity but retains poor specificity and depends substantially on radiologist experience.
- Engineered-feature machine-learning models require substantial domain expertise, producing accuracy that varies with feature quality.
- XmasNet applies end-to-end deep CNN training to 3D multiparametric MRI for prostate cancer diagnosis.
- The study addresses heterogeneous medical-imaging data and small sample sizes through preprocessing and data augmentation.
2. DATA PREPROCESSING AND DATA AUGMENTATION
The preprocessing pipeline standardizes and aligns multiparametric MRI, refines lesion centers, and prepares stratified training and validation samples. Three-dimensional augmentation generates multi-view 2D inputs while retaining lesion information across orientations.
- 341 cases provided DWI, ADC, Ktrans, and T2WI data with lesion labels and locations.
- Images were interpolated to 1mm isotropic resolution and co-registered using transverse T2WI as the reference.
- After removing five low-quality cases, 169 cases with 274 lesions formed training data and 30 cases with 43 lesions formed validation data.
- Augmentation used seven slicing orientations plus in-plane rotation, random shearing, translation, and random mirroring during Caffe training.
- 207144 training samples were prepared as 32×32 lesion-centered ROIs, while test predictions averaged outputs across orientations.
3. DEEP LEARNING MODEL: XMASNET
XmasNet is a convolutional architecture trained separately on multiple input types and combined through a validation-guided ensemble. Its design uses convolutional, normalization, activation, pooling, and fully connected layers.
- XmasNet combines convolutional, batch-normalization, ReLU, max-pooling, and fully connected layers.
- Twenty XmasNet models using different input types were trained, and lesion predictions were formed from weighted model averages.
- A greedy bagging algorithm selected models that maximized ensemble validation AUC during iterative ensemble construction.
- Table 1 organizes each XmasNet layer by patch size, stride, and output size.
- Seven XmasNets received non-zero ensemble weights, including five relatively higher-AUC models and two relatively lower-AUC models.
4. FEATURE ENGINEERING AND DECISION TREE MODEL
The conventional comparison system used engineered intensity and texture features with independently tuned gradient-boosted decision-tree ensembles. Functional and texture features, particularly from Ktrans and ADC, appeared frequently among selected features.
- Each lesion was represented by 87 intensity, statistical, and 3D Haralick texture features.
- 140 XGBoost decision-tree models underwent backward feature selection and hyper-parameter tuning through independent 5-fold cross-validation.
- The top 20 models by cross-validation AUC formed an ensemble using greedy bagging.
- Frequently appearing features were mainly functional or texture features associated with Ktrans and ADC.
5. COMPARISON OF DEEP LEARNNG AND XGBOOST MODELS
XmasNet outperformed XGBoost on both train and test data, while Figure 4 compares their ROC curves and model sensitivities and specificities.
- XmasNet outperformed XGBoost on both train and test data by AUC.Train AUCs were 0.95 versus 0.89, and test AUCs were 0.84 versus 0.80, for XmasNet and XGBoost, respectively.
- 0.84 was the test-set AUC for XmasNet, compared with 0.80 for XGBoost.
- 0.89 was the validation sensitivity and specificity achieved using XmasNet.
- Figure 4 compares ROC curves on validation data and sensitivities and specificities across models.
6. CONCLUSIONS AND FUTURE DIRECTIONS
The study concludes that XmasNet demonstrates strong potential for prostate cancer imaging, while identifying feature combination and more sophisticated architectures as future improvements.
- XmasNet outperformed 69 methods from 33 participating groups and achieved the second-highest AUC in the 2017 PROSTATEx challenge.
- 3D rotation and slicing were used for data augmentation to capture lesion 3D information and avoid overfitting.
- XmasNet could be further improved by combining hand-crafted features and using more sophisticated architectures.
- The deep learning model outperformed the conventional feature-engineering model in both train and test sets.