---
tags: Medical
---
# 論文閱讀: Convolutional neural network based deep-learning architecture for prostate cancer detection on multiparametric magnetic resonance images (2017)

## Introduction
* Prostate cancer (PCa) is the second most common cause of cancer related deaths in men.
* Multiparametric MRI (mpMRI,多參數MRI) is the most accurate imaging method for PCa detection.
* We adopted a network architecture from a state-of-the-art edge detector that takes an image as an input and produces an image probability map.
* Two-fold cross validation along with a receiver operating characteristic (ROC) analysis and free-response ROC (FROC) were used to determine our deep-learning based prostate-CAD’s (CADDL) performance.
* The efficacy was compared to an existing prostate CAD system that is based on hand-crafted features, which was evaluated on the same test-set.

## Materails and Methods
### Patient Cohort
* Consisted of 52 patients with T2W, ADC, and B2000 MR images
* For every patient we also had a prostate contour segmentation acquired using technique [Ref](https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6944225)
* Tumor annotations done by an expert radiologist. Figure 1 shows an example of the original image sequences and the regions of interest.
* The patients in the study presented with lesions(病灶) of various severity(嚴重程度) and location. Biopsy(活體組織切片) results are used to characterize the lesions. Table 1 has all the biopsy points for the study cohort stratified(分層) by zone (peripheral or transition) and Gleason score(6-9), where the higher the score the more severe the lesion.
### Image Pre-processing Step
* CNNs are mostly designed for natural images which usually have three channels. To satisfy this constraint, we generated a three channel RGB image from three MR sequences as shown in Figure 2.
* Usually, the input must be an 8 bit image with intensity values between 0 and 255; however, medical images such as the mpMRIs that we used are 12-16 bit with intensity values extending to upwards of 4000.
* We performed histogram equalization for all images to minimize information loss during compression.
* Figure 3 shows the effects of using a histogram equalizing algorithm
### CNN Architecture
* Adopted the CNN architecture of Holistically-Nested Edge Detection(相關可以參考以下連結)
* [Holistically-Nested Edge Detection](https://hackmd.io/@RinHizakura/rk9cVKHz8)

* The architecture has multiple side outputs as shown in Figure 4.

特別需提到的是，使用的loss function如下。
其中Y+表示被標註為病灶區的部分，Y-為非病灶區的部分，Y=(Y+) + (Y-)。這個loss function設計的主要目的是因為標註是與標註非病灶區的數量不平衡，透過這種設計可以避免unbalance的資料影響預測結果過多。

### Analysis
#### Cross-validation
* The patient cohort was divided equally into two sets which were used for training or testing depending on the fold. Due to the limited number of patients in the study cohort, a validation-set was left out of the experimental setup. Instead, the CAD was evaluated on the test-set at multiple learning stages to characterize its performance over multiple epochs.
#### ROC Analysis
* ROC analysis is performed on 3D probability maps. The 3D probability maps are first formed for each case by taking the 2D probability map outputs from CADDL and stacking them.
* Detection rate is then determined by computing the 90% of the probability scores within each cancerous lesion volume. If this exceeds some threshold, then the cancerous lesion is said to be detected. In other words, if at least 10% of the lesion has relatively high probability, it is considered detected.
* False positive rate is then determined by placing a 3mm x 3mm x 3mm grid on the prostate. Only cells that are inside the prostate are considered. If the 90th percentile of probability scores in a cell exceeds a threshold, then the cell is said to be a false positive.
#### Comparison
* The performance of our CADDL was compared to the published competing CADSVM by [Ref](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4401803/)
* CADSVM uses local binary pattern features extracted from T2WI and B2000 images as features to an SVM classifier.
* In contrast to CADDL, which was trained on a relatively small number of tumor contour segmentations, biopsy point examples from 108 patients were used to train CADSVM. The testing-sets from CADDL were used to compute curves and generate images for CADSVM.

## Results
* In Table 2, the detection rate at 0.1 false-postive rate (FPR) increases with increasing margin until it reaches maximums at epoch 5. All the models have a higher AUC and all except the model at epoch-1 have a higher detection rate at 0.1 FPR than CADSVM.
* The ROC plots for four different models are presented in Figure 6. The first three are for CADDL models at different epochs, and the last curve is for CADSVM
* Figure 8 shows a detection of a transition zone lesion that appears suspicious on all MR sequences –lesions usually have low intensity on T2W and ADC images while appearing bright on B2000.
* CADDL detects it with high confidence while generating no false-positives.
* CADSVM has a true-positive detection as well, but presents false-positive detections visible in the peripheral zone.
* Benign prostatic hyperplasia (BPH) can be erroneously identified as cancerous by both CADs as illustrated in Figure 9. However, in this instance only CADDL manages to generate high probability values for the peripheral zone lesion that cohabits this region of the prostate.
* Some lesions can be difficult to identify for the CADs, such as the small peripheral zone tumor presented in Figure 10. In this example, both CADs incorrectly predict a large tumor in the transition zone where there are no marked regions of interest.
* Lastly, false predictions can also arise from an artifact within the image. Figure 11 presents an example where there is a blurry region in the peripheral zone of the T2W image. CADDL incorrectly predicts suspicion in this region with high probability while CADSVM remains largely unaffected by the low quality image.
## Discussion
### problem and limitation
* Due to a limited data-set size, the CAD achieves its optimal performance after a small number of epochs
* A small data-set makes it likely that the training-set excludes some prostate lesion types that may be found in the test-set. Since CNNs learn important features from the data, the CAD will be unable to detect such lesion types because of the lack of corresponding features.
* A small study cohort is one limitation in our study. We investigate CADDL performance over multiple epochs instead of finding the best performing model using a validation-set.
* A second limitation in our study is the loss of information during image compression to satisfy the constraints of using publicly available CNN architectures. Even though we apply a histogram equalization algorithm to convert the medical images from a 12-16 bit to 8 bit images, there remains a small amount of information loss.
### future work
* As a future study, we plan to investigate a weakly supervised learning scheme that can use the more ubiquitous database types containing biopsy points and histopathology images for training.
* Acquiring a large enough dataset with ground-truth tumor annotations is difficult because it required the expertise of experienced healthcare providers, who are preoccupied with patient care.
* Furthermore, our chosen CNN architecture can be modified for multi-label classification, which will allow us to identify lesions by severity in addition to detecting them.
* Lastly, we will attempt to develop an automatic segmentation tool for tumor size prediction.
## Conclusion
* Conclude that deep convolutional neural networks (DCNNs) have the potential to improve prostate cancerdetection on multiparametric MR images (mpMRI).
* We demonstrated that our proposed DCNN-based prostate CAD (CADDL) is capable of learning lesion discriminating features from the given mpMR images.
* CADDL further achieved promising results when compared to a published prostate CAD and showed potential for clinical application.
* In the future, we hope to develop a CAD that has many more applications in addition to tumor detection such as severity and tumor size prediction.