Source-linked AI summary
Detection-aided liver lesion segmentation using deep learning
Miriam Bellver, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Xavier Giro-i-Nieto, Jordi Torres, Luc Van Gool
TL;DR
The paper addresses automatic liver-lesion segmentation in CT, where low contrast, lesion-size variability, and noise make robust segmentation difficult. It combines cascaded DRIU-based segmentation with liver-focused training, context slices, lesion detection, and 3D-CRF postprocessing; the detector improves results by filtering false positives.
Problem
Robust automatic liver-lesion segmentation is difficult because CT images contain low contrast, variable lesion sizes, and noise, despite its clinical value.
Source: §1 · p. 1Method
The method cascades liver and lesion segmentation, uses DRIU with liver-focused weighted training and context slices, adds a lesion detector, and applies 3D-CRF postprocessing.
Source: §2 · p. 2, §2.1 · p. 2, §2.1.1 · p. 2Results
Takeaways & Limitations
Combining pixel-wise segmentation with detector-based global decisions provides a beneficial detection-aided pipeline for localizing liver lesions.
Source: §4 · p. 4
Abstract
from arXiv · showhide
A fully automatic technique for segmenting the liver and localizing its unhealthy tissues is a convenient tool in order to diagnose hepatic diseases and assess the response to the according treatments. In this work we propose a method to segment the liver and its lesions from Computed Tomography (CT) scans using Convolutional Neural Networks (CNNs), that have proven good results in a variety of computer vision tasks, including medical imaging. The network that segments the lesions consists of a cascaded architecture, which first focuses on the region of the liver in order to segment the lesions on it. Moreover, we train a detector to localize the lesions, and mask the results of the segmentation network with the positive detections. The segmentation architecture is based on DRIU, a Fully Convolutional Network (FCN) with side outputs that work on feature maps of different resolutions, to finally benefit from the multi-scale information learned by different stages of the network. The main contribution of this work is the use of a detector to localize the lesions, which we show to be beneficial to remove false positives triggered by the segmentation network. Source code and models are available at https://imatge-upc.github.io/liverseg-2017-nipsws/ .
1 Introduction
Liver-lesion segmentation supports cancer diagnosis and treatment-response assessment, but remains difficult because of low contrast, variable lesion size, and CT noise. The paper builds on CNN-based methods to address this open problem.
- Low contrast, variable lesion size, and CT noise make robust liver-lesion segmentation an open problem.Source: §1 · p. 1
- Automatic segmentation could reduce the subjectivity and time required by manual interpretation of CT and MRI images.Source: §1 · p. 1
- CNN-based methods have demonstrated robustness to these challenges and achieved state-of-the-art performance for this task [3] [5] [1].Source: §1 · p. 1
2 Detection-aided liver and its lesions segmentation using deep learning
The method cascades liver and lesion segmentation, augments DRIU with imbalance-aware and liver-focused training, adds context slices and lesion detection, and applies 3D-CRF postprocessing.
2 Detection-aided liver and its lesions segmentation using deep learning
- The pipeline first segments the liver, then crops its region for lesion segmentation, and retains lesion pixels supported by the detector.Source: Figure 1 · p. 2, §2 · p. 2, §2.2 · p. 3
2.1 Segmentation network
- DRIU-based networks combine supervised side outputs from multiple feature-map resolutions to produce multi-scale liver and lesion segmentations.Source: §2.1 · p. 2
2.1.1 Loss objective
- Weighted BCE addresses lesion-pixel imbalance, while liver-only gradient propagation restricts learning to pixels that can belong to lesions.Source: §2.1.1 · p. 2, §2.1.2 · p. 3
2.1.3 Using context slices to exploit 3D information
- Three consecutive CT slices are input simultaneously to exploit volume redundancy, while only the central slice is retained at test time.Source: §2.1.3 · p. 3
2.2 Lesion detection network
- The detector supplies a global healthy/unhealthy decision that removes segmentation false positives when it disagrees with pixel-wise lesion predictions.Source: §2.2 · p. 3, Figure 2 · p. 3
2.3 Postprocessing
- A 3D fully connected CRF performs final postprocessing using spatial coherence and input-volume intensity appearance.Source: §2.3 · pp. 3–4
3 Experimental validation
Experiments on the LiTS Challenge test compare four submissions, showing improvements from the cascaded segmentation configuration and from adding the detector and 3D-CRF.
- The LiTS Challenge test evaluates four submissions and reports Dice scores for the baseline, enhanced segmentation, detector, and 3D-CRF configurations.Source: §3 · p. 4
- The 3-i/o + BP-in-liver segmentation configuration performs considerably better than the cascaded baseline.Source: §3 · p. 4
- The table also reports improvements achieved by adding the lesion detector and 3D-CRF on the official LiTS testing set.Source: §3 · p. 4
4 Conclusions
The proposed pipeline combines cascaded liver and lesion segmentation with lesion detection to improve automatic liver-lesion analysis. Its conclusions emphasize liver-focused learning and the complementary strengths of pixel-wise segmentation and patch-level detection.
- The proposed algorithm uses two cascaded segmentation networks and a lesion detector, while exploiting volume information and liver predictions.Source: §4 · p. 4
- The method learns lesion representations only from liver pixels, focusing training on relevant samples and improving the positive-negative pixel imbalance.The authors relate this liver-focused strategy to attention mechanisms because it selects the liver as the relevant learning region.Source: §4 · p. 4
- Combining pixel-wise segmentation with lesion detection yields a better overall result on the LiTS Challenge database.Segmentation provides fine localization, while detection contributes stronger generalization and suppresses false positives through more global decisions.Source: §4 · p. 4