Author: Ilias Gatos, Stavros Grigoriadis, George C. Kagadis, Maria Karamesini, Paraskevi Katsakiori, Dimitris N. Mihailidis, Stavros Spiliopoulos, Efstratios Syrmas, Ioannis Theotokas, Stavros Tsantis, Pavlos Zoumpoulis 👨🔬
Affiliation: Diagnostic Echotomography, University of Pennsylvania, University of Athens, University of Patras 🌍
Purpose: To detect prostate lesions on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) images which is a particularly difficult task due to the heterogeneous and inconsistent representation of the pixels surrounding the prostate. Large-scale manual annotations are time-consuming. This study aims to evaluate the agreement between the Edge-Driven Modified Fuzzy C-Means (EDM-FCM) semi-automatic segmentation algorithm and manual segmentation by an expert radiologist in DCE-MRI prostate cancer lesions’ images.
Methods: DCE-MRI images of 65 patients with biopsy-confirmed prostate cancer were used. One DCE image from every patient was selected during the wash-in phase. Each image contained one lesion. Lesions were manually delineated by an expert radiologist. The algorithm comprised of the following steps: (a) manual selection of a bounding box containing the lesion, (b) calculation of the edge indicator function (EIF) of the selected ROI using continuous wavelet transform with the Mexican Hat filter, (c) region clustering via EDM-FCM algorithm using mean intensity values between consecutive edges as input. Agreement between automatic segmentation and manual contours was evaluated with the Jaccard Index, Dice Coefficient, and Hausdorff distance indices.
Results: Jaccard Index value ranged between 0.24 and 0.71 with an average value of 0.48. Dice Coefficient Index values ranged between 0.36 and 0.83 with an average value of 0.64. Hausdorff distance value (mm) ranged between 2.14 and 13.91 with an average value of 5.18. The results indicate moderate agreement over the methods used.
Conclusion: EDM-FCM showed a level of agreement with manual segmentation that is classified as moderate. However, the results are inferior compared to those reported in the literature, suggesting that the use of DCE as a unique sequence limits the detection accuracy. More studies are deemed necessary to prove the potential of the current approach.