MILD-Net: Minimal Information Loss Dilated Network for Gland Instance Segmentation in Colon Histology Images (1806.01963v4)

Abstract: The analysis of glandular morphology within colon histopathology images is an important step in determining the grade of colon cancer. Despite the importance of this task, manual segmentation is laborious, time-consuming and can suffer from subjectivity among pathologists. The rise of computational pathology has led to the development of automated methods for gland segmentation that aim to overcome the challenges of manual segmentation. However, this task is non-trivial due to the large variability in glandular appearance and the difficulty in differentiating between certain glandular and non-glandular histological structures. Furthermore, a measure of uncertainty is essential for diagnostic decision making. To address these challenges, we propose a fully convolutional neural network that counters the loss of information caused by max-pooling by re-introducing the original image at multiple points within the network. We also use atrous spatial pyramid pooling with varying dilation rates for preserving the resolution and multi-level aggregation. To incorporate uncertainty, we introduce random transformations during test time for an enhanced segmentation result that simultaneously generates an uncertainty map, highlighting areas of ambiguity. We show that this map can be used to define a metric for disregarding predictions with high uncertainty. The proposed network achieves state-of-the-art performance on the GlaS challenge dataset and on a second independent colorectal adenocarcinoma dataset. In addition, we perform gland instance segmentation on whole-slide images from two further datasets to highlight the generalisability of our method. As an extension, we introduce MILD-Net+ for simultaneous gland and lumen segmentation, to increase the diagnostic power of the network.

• The paper introduces MILD-Net, a novel convolutional network that preserves fine details using minimal information loss units and dilated convolutions.
• It utilizes atrous spatial pyramid pooling and random transformation sampling to effectively handle multi-scale features and quantify segmentation uncertainty.
• MILD-Net achieves state-of-the-art performance on colon gland segmentation benchmarks and demonstrates robust results across varied datasets and whole-slide images.

Overview of MILD-Net: Minimal Information Loss Dilated Network for Gland Instance Segmentation

This paper introduces MILD-Net, a novel fully convolutional neural network designed to address the complexities of gland instance segmentation in colon histology images. The primary motivation behind this research is the necessity for automated segmentation methods in pathology due to the labor-intensive and subjective nature of manual gland analysis, crucial for determining colorectal cancer grades. The variability in glandular morphology and the distinction between glandular and non-glandular structures pose significant challenges, which MILD-Net is engineered to overcome.

The architecture of MILD-Net innovatively counters information loss, typically a consequence of max-pooling, by reincorporating original image details at multiple points within the network. This approach is bolstered by atrous spatial pyramid pooling (ASPP) to maintain resolution and handle multi-scale features effectively. Additionally, MILD-Net integrates uncertainty quantification through random transformation sampling (RTS), generating an uncertainty map that highlights ambiguous regions within the segmentation output. This uncertainty map is pivotal for discarding predictions with high uncertainty, enhancing reliability in diagnostic applications.

Key Contributions and Results

1. Innovative Network Architecture: MILD-Net utilizes minimal information loss (MIL) units combined with dilated convolutions to maximize detail retention throughout feature extraction. The network's architecture is explicitly designed to preserve important details necessary for precise gland boundary delineation.
2. Multi-scale Feature Handling: Through ASPP, the network effectively aggregates features at multiple scales, crucial for handling glands of varying shapes and sizes, particularly in challenging high-grade cancer cases.
3. Uncertainty Mapping: By applying RTS during test-time, MILD-Net not only augments segmentation accuracy but also provides a quantitative measure of uncertainty across predictions. This is critical in clinical contexts where diagnostic confidence is as important as accuracy.
4. State-of-the-art Performance: The model achieves superior results on the MICCAI 2015 Gland Segmentation (GlaS) challenge and an independent colorectal adenocarcinoma dataset, surpassing existing methods in several key evaluation metrics, including F1 score and object-level dice.
5. Generalizability: The network demonstrates robustness by maintaining high performance across different datasets and even in whole-slide image (WSI) processing, underscoring its potential for broader applicability in clinical settings.
6. MILD-Net$^+$ Extension: An extension, MILD-Net$^+$, introduces simultaneous segmentation of glands and gland lumens, enhancing the network's utility by providing additional morphological details essential for accurate cancer grading.

Practical and Theoretical Implications

Practically, MILD-Net can significantly reduce the workload of pathologists by automating the tedious task of gland segmentation, thus allowing them to focus on higher-level interpretative work. Its ability to quantify uncertainty also adds a layer of reliability, ensuring that only confident predictions are considered in diagnostic proceedings. Theoretically, the framework offers a new perspective on preserving spatial information in deep networks, a crucial factor for accurate segmentation tasks.

Speculation on Future Developments

Future developments might focus on optimizing MILD-Net for faster WSI processing, potentially revolutionizing digital pathology workflows. Furthermore, applying MILD-Net to other histopathological tasks across different cancer types could broaden its impact. Finally, combining MILD-Net with advanced machine learning techniques, such as active learning, could minimize the dependency on extensive labeled datasets, thus facilitating a smoother integration into clinical practice.

In conclusion, MILD-Net presents a significant advancement in the automation of histopathological image analysis, addressing both current challenges and setting a foundation for future enhancements in the field of computational pathology.