Computer-Aided Detection and diagnosis for prostate cancer based on mono and multi-parametric MRI: A review
Introduction
During the last century, physicists have focused on constantly innovating in terms of imaging techniques assisting radiologists to improve cancer detection and diagnosis. However, human diagnosis still suffers from low repeatability, synonymous with erroneous detection or interpretations of abnormalities throughout clinical decisions [1], [2]. These errors are driven by two majors causes [1]: observer limitations (e.g., constrained human visual perception, fatigue or distraction) and the complexity of the clinical cases themselves, for instance due to unbalanced data (number of healthy cases more abundant than malignant cases) or overlapping structures.
Computer vision has given rise to many promising solutions, but, instead of focusing on fully automatic computerized systems, researchers have aimed at providing computer image analysis techniques to aid radiologists in their clinical decisions [1]. In fact, these investigations brought about both concepts of Computer-Aided Detection (CADe) and Computer-Aided Diagnosis (CADx) grouped under the acronym CAD. Since those first steps, evidence has shown that CAD systems enhance the diagnosis performance of radiologists. Chan et al. reported a significant 4% improvement in breast cancer detection [3], which has been confirmed in later studies [4]. Similar conclusions were drawn in the case of lung nodule detection [5], colon cancer [6] and Prostate Cancer CaP as well [2]. Chan et al. also hypothesized that CAD systems will be even more efficient assisting inexperienced radiologists than senior radiologists [3]. This hypothesis was tested by Hambrock et al. [2] and was confirmed in the case of CaP detection. In this particular study, inexperienced radiologists obtained equivalent performance to senior radiologists, both using CAD whereas the accuracy of their diagnosis was significantly poorer without CAD׳s help.
In contradiction with the aforementioned statement, CAD for CaP is a young technology due to the fact that it is based on Magnetic Resonance Imaging (MRI) [7]. Four distinct MRI modalities are employed in CaP diagnosis which were mainly developed after the mid-1990s: (i) T2 Weighted (T2-W) MRI [8], (ii) Dynamic Contrast-Enhanced (DCE) MRI [9], (iii) Magnetic Resonance Spectroscopy Imaging (MRSI) [10] and (iv) Diffusion Weighted (DW) MRI [11]. In addition, the increase of magnetic field strength (from 1.5 to 3 Tesla) and the development of endorectal coils, both improved image spatial resolution [12] needed to perform more accurate diagnosis. It is for this matter that the development of CAD for CaP is still lagging behind the other fields stated above.
The first study on CAD for MRI was published in 2003 by Chan et al. [13]. Despite this, no less than 50 studies have been reviewed for this survey since that seminal work. To the best of our knowledge, there is no fully detailed review in the literature regarding the advancement of CAD systems devoted specifically to CaP detection and diagnosis. Only the recent work of [14] briefly covers a reduced number of current research works on CAD for CaP (a total of 70 references), but it does not provide a detailed description of any of the steps of a CAD system as it is done here with more than 200 references. Moreover, a new categorization is proposed in this work taking the clinical and technical aspects of a CAD system into account. Thus, the aim of this survey is threefold: (i) provide an overview and categorization of the developed CAD systems for CaP detection and diagnosis based on MRI modalities (ii) assess the different works and (iii) point out avenues for future directions.
We also would like to emphasize the fact that this study will review the details of the computer vision aspects of the different studies. Thus, this survey is more intended for a medical imaging audience rather than a purely experimented clinical audience.
As discussed further in Section 2.2.3, we identified and characterized a common framework regarding the CAD systems. Stages involved in CAD work-flow can be categorized into three distinctive processes: (i) image regularization, (ii) CADe, (iii) CADx (see Fig. 1). Image regularization focuses on formatting the data while CADe and CADx allow us to detect possible lesions and distinguish malignant from non-malignant tumours, respectively.
This paper is organized as follows: Section 2 deals with general information about human prostate and background about CaP. Methods regarding CaP screening and imaging techniques used are also presented as well as an introduction to the CAD framework. 3 Image regularization framework, 4 CADe–CADx review techniques used in different steps involved in a CAD work-flow which will be our main contribution. Image regularization framework including pre-processing (Section 3.1), segmentation (Section 3.2) and registration (Section 3.3) will be covered as well as CADe and CADx strategies (Section 4) identified. Results and discussion are reported in Section 5 followed by a concluding section. Deriving from this discussion, we make available to the research community a public online dataset aiming at overcoming some of the drawbacks found when evaluating research in this field.
Section snippets
Background
This section provides an overview of CaP as well as its detection and diagnosis. MRI plays an important role in improving the current strategy and a more detailed description of MRI modalities is given. Furthermore, a discussion regarding the aim of CAD systems is also given.
Image regularization framework
This section provides a review of the methods used in CADs for CaP in order to regularize input images. We start with pre-processing methods presented in Section 3.1, focusing mainly on the reduction of noise level and artifacts as well as standardization of SI. 3.2 Segmentation, 3.3 Registration will be dedicated to segmentation methods, so that later methods only operate on the segmented prostate, and registration to align segmented images from different MRI-modalities in the same reference
CADe: ROIs detection/selection
As discussed in the introduction and shown in Fig. 1, the image classification framework is often composed of a CADe and a CADx. In this section, we will focus on studies embedding a CADe in their framework. Two approaches are considered to define a CADe (see Table 6): (i) voxel-based delineation and (ii) lesion segmentation. The first strategy, which concerns the majority of the studies reviewed (see Table 6), is in fact linked to the nature of the classification framework [114], [115], [163],
Results reported
As discussed previously in Section 4.4.3, different metrics have been used to report results. A comparison of the different methods reviewed is given depending on the metric used in field of research and also the type of MRI scanner used (cf., 1.5 versus 3.0 Tesla). For each field, the best performances obtained in each study were reported in these figures. The results given in terms of AUC-ROC are depicted in Fig. 9. The results vary between 71% and 97% for some experiments with a 1.5 Tesla MRI
Conclusion
This review has presented an overview and classification of the research related to CAD development for CaP using multi-parametric MRI data. We aimed at providing background information regarding multi-parametric MRI imaging techniques and a description of the work-flow in the different CAD stages. The methods used in the literature for each of these stages have been reviewed along with the available results of the CAD systems. Moreover, insight discussions and possible future research
Conflict of interest statement
The authors declare no conflict of interest.
Acknowledgments
G. Lemaître was supported by the Generalitat de Catalunya (Grant no. FI-DGR2012) and partly by the Mediterranean Office for Youth (Grant no. 2011/018/06).
We would like to acknowledge Sharad Nagappa for all the discussions involved and his precious advices and corrections regarding the reduction of this entire manuscript.
References (218)
- et al.
Histological and clinical findings in 896 consecutive prostates treated only with radical retropubic prostatectomyepidemiologic significance of annual changes
J. Urol.
(1998) - et al.
Central zone carcinoma of the prostate glanda distinct tumor type with poor prognostic features
J. Urol.
(2008) - et al.
Mortality results from the Göteborg randomised population-based prostate-cancer screening trial
Lancet Oncol.
(2010) - et al.
Prebiopsy magnetic resonance imaging and prostate cancer detectioncomparison of random and targeted biopsies
J. Urol.
(2013) - et al.
In vivo quantification of citrate concentration and water T2 relaxation time of the pathologic prostate gland using 1H MRS and MRI
Magn. Reson. Imaging
(1997) - et al.
The role of choline in prostate cancer
Clin. Biochem.
(2012) - et al.
Proton MR spectroscopy of prostatic tissue focused on the detection of spermine, a possible biomarker of malignant behavior in prostate cancer
MAGMA
(2000) - et al.
MRI denoising using non-local means
Med. Image Anal.
(2008) - et al.
A survey on the magnetic resonance image denoising methods
Biomed. Signal Process. Control
(2014) - et al.
Anniversary paperhistory and status of CAD and quantitative image analysis: the role of Medical Physics and AAPM
Med. Phys.
(2008)