Abstract
Magnetic resonance imaging (MRI) is essential for prostate cancer diagnosis and treatment planning. However, manual segmentation of the prostate and its internal zones is time-consuming and subjective. We propose Dense-2 U-net, a hybrid architecture combining DenseNet's feature reuse with U-net's localization capabilities. Evaluated on 188 patient datasets, our model achieved a Dice score of 92.1% for the whole prostate, outperforming the standard U-net while requiring significantly fewer parameters (15.6M vs. 33M). Notably, the model demonstrates robustness when trained on weak annotations, often producing results more anatomically accurate than the initial noisy labels.
Introduction
Accurate segmentation of the prostate gland and its internal regions—the Central Zone (CZ) and Peripheral Zone (PZ)—is vital for diagnosing prostate cancer, planning surgery, and monitoring treatment. Manual segmentation is a significant burden for radiologists, characterized by high subjectivity and inter-observer variation. Automating this task is challenging due to the prostate's "fuzzy" boundaries, tissue heterogeneity, and significant inter-individual shape variability. This study introduces a novel neural network designed to address these challenges through efficient feature propagation and reuse.
Methods and Architecture
The proposed Dense-X U-net architecture integrates "Dense Blocks" into the traditional U-net framework. In a Dense Block, each layer receives feature maps from all preceding layers as input. This feature map concatenation ensures that the network retains critical spatial information that might otherwise be lost during successive convolutions.
To manage computational complexity, transitional layers are used to compress feature maps. This design allows the Dense-2 U-net to operate with only 15.6 million parameters, less than half of the 33 million parameters used by a standard U-net. The researchers found that standard Cross-entropy loss provided the most stable training performance for this specific multi-zone segmentation task.
Datasets and Ground Truth
The study utilized the PROSTATEx challenge dataset, comprising T2-weighted MR images from 188 patients. The model was trained on 141 patient datasets (2,982 slices) and tested on 47 (912 slices). A key component of the study was testing the model's ability to learn from both "coarse" (weakly annotated) and "fine" (accurately revised) ground truth masks.
Key Results and Insights
The Dense-2 U-net achieved superior quantitative results compared to the classical U-net across all target regions:
- Whole Prostate (PR): 92.1% Dice score (vs. 90.7% for U-net).
- Central Zone (CZ): 89.5% Dice score (vs. 89.1% for U-net).
- Peripheral Zone (PZ): 78.1% Dice score (vs. 75.0% for U-net).
Beyond raw metrics, the Dense-2 model showed a remarkable ability to handle difficult "edge cases," such as distinguishing the prostate from the bladder in patients with Benign Prostatic Hyperplasia (BPH). Furthermore, the network demonstrated a "self-correcting" property: when trained on weak labels, it often produced predictions that were more anatomically consistent than the noisy training data itself. The model's performance on the whole prostate (92.1%) even surpassed the intra-observer reliability of a human radiologist (90.0%).
Limitations and Future Work
While highly effective, the model faces challenges at the prostate's apex and base, where boundary ambiguity is highest and training data is less abundant. Additionally, the current 2D slice-by-slice approach, while efficient, does not fully exploit the 3D volumetric context of the MRI data. Future research will focus on improving segmentation in these difficult anatomical regions and exploring 3D spatial context to further refine accuracy.
Summary updated