Comparison of Machine and Deep Learning models for automatic segmentation of prostate cancers on multiparametric MRI

Abstract

Multiparametric (mp) magnetic resonance imaging (MRI) represents a robust tool for detecting prostate cancers (PCa). However, its interpretation requires skilled and specialized staff, and large investments of resources and time. To deal with this problem different artificial intelligence algorithms, based on Machine Learning (ML) and Deep Learning (DL), have been proposed and have been demonstrated useful to detect and characterize PCa. In this paper, we present a fully automated computer-aided diagnosis (CAD) system that utilizes either ML or DL techniques to segment PCa and we compared the results in terms of number of False Negative (FN) and False Positives (FPs) findings and accuracy of the segmentation masks. We present a DL model with two different input configurations: 2-channel and 3-channel. According to our results, DL techniques greatly decrease the volume of FPs and the number of FN compared to ML techniques, especially using the 3-channel model. Indeed, on the validation set, the number of FNs obtained by the DL model is lower than that by ML (respectively 7 and 11), while the median volume of FPs voxels decreased from 1077 <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{IQR}=362- 3787$</tex> to 518 <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{IQR}=170-1049$</tex> . The results obtained from this system could have a fairly obvious improvement by increasing the validation set, however preliminary results are encouraging and could be a strong contribution for personalized medicine.