A comparative analysis of Convolutional Neural Network architectures for classifying the degree of risk of prostate lesions from unimodal or bimodal mpMRI images

Abstract

This work presents a comparative analysis of five convolutional neural network (CNN) architectures using multiparametric magnetic resonance imaging
(mpMRI) for the classification of tissues with the presence of prostate cancer lesions. SPIE-AAPM-NCI Prostate MR Classification Challenge was used as a
training and validation data set, which consists of 344 cases with magnetic resonance images from the modalities: T2W (Weighted T2), ADC (Apparent Diffusion Coefficient), and Ktrans (preprocessed images from the DCE -Dynamic Enhanced Contrast- modality), from which three subsets of data from a single independent modality were used (unimodal): T2W, ADC and Ktrans, and two subsets of data combining two modalities (bimodal): Ktrans-ADC and Ktrans-T2W, for comparison and analysis. From the Gleason scale (Gleason score - GS) and the ISUP grade (International Society of Urologic Pathologists), which are used to measure the degree of aggressiveness of prostate cancer, two levels of aggressiveness were established: Low and High. The Low class is those lesions with GS = 6, and the High class, those lesions with the GS value > 7. The experimental results show a superior performance with Ktrans modality images in the first 4 architectures obtaining a maximum AUC value (area under ROC curve) of 0.71 ± 0.127. However, the fifth LetNet inspired architecture combining two mpMRI modalities, Ktrans-T2W, obtains an AUC of 0.72 ± 0.058, which slightly suggests that although the Ktrans modality is the
most relevant, its combination with T2W could improve diagnostic accuracy.