Abstract
Convolutional neural networks (CNNs) require large amounts of data for training, beyond what can be acquired for current radiomics models. We hypothesize that deep entropy features (DEFs) derived from existing CNNs can be applied to MRI images of prostate cancers (PCa) to reliably predict the Gleason score (GS) of PCa lesions. In this study, we analyzed 112 lesions acquired from 99 PCa patients, either pre-biopsy or pre-treatment, their associated GS, and multi-parametric MRI (mpMRI) sequences. Our approach is based on the extraction of DEF features produced in individual layers of 9 pre-trained CNN models. Among the 9 pre-trained CNNs, the NASNet-mobile architecture offered the features most correlated to GS ($\rho = 0.47$; p < 0.05). The RF classifier discerned GS of 6, 3+4, 4+3, 8 and ≥9 with an AUC (%) of 80.08, 85.77, 97.30, 98.20, and 86.51, respectively. Our results suggest that the DEFs can be used to differentiate GS of PCa lesions with high accuracy, potentially improving diagnosis and individualizing patient care.
Introduction
The Gleason Score (GS) is the clinical gold standard for grading prostate cancer (PCa) aggressiveness. However, determining this score traditionally requires invasive needle biopsies, which carry risks of infection and are subject to sampling errors. While multi-parametric MRI (mpMRI) is a powerful non-invasive alternative, interpreting these images is complex. Standard radiomics often leads to overfitting on small datasets, while standard deep learning typically requires thousands of labeled images. This paper proposes a middle ground: "Deep Entropy Features" (DEFs) that leverage the representational power of deep learning with the statistical stability of information theory.
Methodology and Feature Extraction
The researchers developed a transfer learning pipeline that bypasses the need for massive medical training data. They utilized 9 well-known CNN architectures (such as AlexNet, ResNet101, and NASNet-mobile) pre-trained on natural images from ImageNet.
Instead of using the final classification layer, the team extracted feature maps from every internal layer of these networks. They then calculated the Shannon Entropy of these maps to serve as a mathematical proxy for tissue heterogeneity and texture complexity. This process generated 7,857 unique features from T2-weighted (T2-WI), Apparent Diffusion Coefficient (ADC), and Dynamic Contrast Enhancement (DCE) sequences.
Experimental Results
The study utilized the PROSTATEx-2 dataset, consisting of 112 lesions from 99 patients. A Random Forest (RF) classifier was trained on the DEFs to distinguish between different Gleason Grade Groups.
The results demonstrated exceptional performance, particularly for high-grade cancer. The model achieved a near-perfect AUC of 98.20% for detecting GS 8 and 97.30% for GS 4+3. Interestingly, the NASNet-mobile architecture provided the features most correlated with Gleason Scores ($\rho = 0.47$), suggesting that mobile-optimized architectures may be particularly efficient at capturing local textures relevant to medical imaging.
Conclusion and Future Directions
Deep Entropy Features offer a robust, non-invasive method for predicting prostate cancer aggressiveness from mpMRI scans. By utilizing pre-trained networks, this approach achieves high accuracy without the need for the massive datasets usually required by deep learning. Future research will focus on validating these findings in larger, multi-center prospective studies and exploring the fusion of DEFs with traditional clinical data like PSA levels to further enhance diagnostic precision.
Summary updated