Application of multivariate image analysis techniques (SMB-PLS) to grade prostate tumor aggressiveness from multiparametric MR images

Congress:

ECR 2019

Poster Number:

C-0340

Type:

Scientific Exhibit

Keywords:

Genital / Reproductive system male, Oncology, Computer applications, MR-Diffusion/Perfusion, MR, CAD, Computer Applications-Detection, diagnosis, Statistics, Cancer, Neoplasia, Tissue characterisation

Authors:

E. Aguado-Sarrió, R. Sanz-Requena, R. perez-martinez, L. Marti-Bonmati, J. M. Prats-Montalbán, A. Ferrer-Riquelme; Valencia/ES

DOI:

10.26044/ecr2019/C-0340

DOI-Link:

https://dx.doi.org/10.26044/ecr2019/C-0340

Fig. 1: Figure 1. Parameter blocks and models. The number of variables is obtained...

Fig. 2: Figure 2. (Top) Behaviors obtained from MCR-ALS perfusion model, representing...

Fig. 3: Figure 3. Sequential block diagram procedure of the SMB-PLS iterative method.

Fig. 4: Figure 4. Results for the top-5 group of statistical models after applying...

Fig. 5: Figure 5. Results for the top-5 group of statistical models after applying...

Aims and objectives

To define prostate tumor aggressiveness combining MR-derived and clinical imaging biomarkers by applying statistical image analysis techniques (sequential multi-block partial least squares, SMB-PLS).

Methods and materials

Fifty patients with histologically-confirmed peripheral prostate tumors (15 Gleason≤6, 35 Gleason>6) underwent standard 3T MRI (Philips Achieva), including T2w, DCE and DWI sequences. Lesions were manually segmented on the highest diffusion b-value images. Quantitative parameters were calculated from the tumor ROI using both parametric and non-parametric models (figure 1). Parametric model parameters were obtained from pharmacokinetic models (1st and 2nd generation) and exponential models (IVIM), the T2w (muscle normalized) is also included as an additional parameter. Non-parametric model parameters were obtained using a statistical latent...

Results

The best performing models showed sensitivity and specificity above 90% (figure 4). They included both DWI blocks (DWI-MCR and model-based DWI) as the most relevant. Then, the DCE and T2w blocks were also included, but provided less significant information to improve the classification. The Q2 reached values of 60% and 55% in average for the best group of models (figure 5), which are extremely high and are reflected in the values of sensitivity/specificity. The most relevant blocks of information were the MCR-based models (2 and...

Conclusion

Sequential latent variable classification models such as SMB-PLS can provide high values of specificity and sensitivity for quantitative MRI-based assessment of prostate tumor aggressiveness. New latent variable MCR-based models provided the most important parameters for describing and discriminating tumor aggressiveness.

Personal information

Contact information: [email protected]

References

Lauzon-Gauthier J, Duschesne C. A new multiblock PLS algorithm including a sequential pathway. In: EuroPact 2014, May 6-9, 2014, Barcelona, Spain. 130. Azari K, Lauzon-Gauthier J, Tessier C, Duchesne C. Establishing multivariate specification regions for raw materials using SMB-PLS, In: “Proceedings of the 9th IFAC symposium on advanced control of chemical processes”, ADCHEM 2015, Whistler, BC, Canada, 2015, 1132-37. Lauzon-Gauthier J, Manolescu P, Duchesne C. The sequential multi-block PLS algorithm (SMB-PLS): Comparison of performance and interpretability. Chemometrics and intelligent laboratory systems. 2018; 180: 72-83.