Abstract
Automatic kidney tumor segmentation from CT scans is an essential step for computer-aided diagnosis of cancer. In this paper, we present an improved H-DenseUNet for kidney and tumor segmentation. Specifically, we first train the DenseUNet and then fine tune the network with the 3D counterpart. To further increase the performance, we employ both cross-entropy and dice loss. We evaluate our method on the 2019 MICCAI kidney and tumor segmentation challenge. We split the training dataset of the challenge to 200 training set and 10 validation set. On the validation set, our method achieves 97.0% (Dice) for kidney segmentation and 67.2% (Dice) for tumor segmentation. This model is submitted to the challenge for final performance evaluation on the test dataset.
Introduction
Kidney cancer is among the top ten most common cancers, yet diagnosing and planning surgical interventions remains a significant challenge. Tumors exhibit extreme variability in size, shape, and location, often lacking clear boundaries against healthy tissue. While Computed Tomography (CT) is the gold standard for visualization, manual segmentation is time-consuming and prone to human error. This paper addresses the need for automatic semantic segmentation by adapting the H-DenseUNet architecture for the 2019 Kidney and Tumor Segmentation (KiTS19) challenge.
Methods
The authors utilize an improved version of H-DenseUNet, a hybrid architecture designed to combine the computational efficiency of 2D networks with the volumetric context of 3D models.
Two-Stage Training
The training process follows a two-stage pipeline:
- 2D DenseUNet Training: The model first learns high-resolution intra-slice features using a 2D DenseUNet.
- 3D Fine-tuning: The network is then fine-tuned with its 3D counterpart to capture spatial relationships across horizontal CT slices.
Optimization and Loss Functions
To improve convergence over standard Stochastic Gradient Descent, the Adam optimizer with a decaying learning rate was employed. To handle class imbalance—where tumors occupy a small fraction of the scan—the researchers used a dual loss function consisting of Weighted Cross-Entropy and Dice Loss.
Preprocessing and Post-processing
Preprocessing involved truncating Hounsfield Units (HU) to a range of [-200, 250] to isolate relevant soft tissue. Post-processing logic was applied to ensure anatomical plausibility:
- Selecting only the two largest connected regions as kidneys.
- Removing tumor predictions located outside the predicted kidney boundaries.
- Filling holes in the final 3D volumetric prediction.
Datasets
The study utilizes the MICCAI 2019 KiTS19 challenge dataset, comprising 300 unique CT scans. For this work, the 210 scans with available ground-truth masks were split into a training set of 200 and a validation set of 10.
Results
The model's performance was measured using the Dice coefficient, representing spatial overlap. On the validation set, the method achieved 97.0% Dice for kidney segmentation and 67.2% Dice for tumor segmentation.
Qualitative analysis shows that the model accurately captures kidney boundaries, while tumor segmentation remains more challenging due to morphological variability. The total processing time per patient ranges from 30 to 200 seconds, suggesting feasibility for clinical workflows.
Discussion and Conclusion
The "hybrid advantage" of H-DenseUNet lies in its ability to achieve superior spatial awareness without the massive computational overhead of training a pure 3D network from scratch. The inclusion of anatomical post-processing steps highlights that deep learning models benefit significantly from basic biological priors.
Despite these successes, a performance gap remains between kidney and tumor segmentation. Future work should focus on improving feature extraction for small or low-contrast lesions. Additionally, while the model performed well on the internal validation set, evaluation on the 90 held-out test scans is necessary for final performance assessment. The hardware requirements (4 NVIDIA Titan Xp GPUs) also suggest that further optimization for lower-end hardware could broaden the tool's accessibility.
Summary updated