Source-linked AI summary
Liver lesion segmentation informed by joint liver segmentation
Eugene Vorontsov, An Tang, Chris Pal, Samuel Kadoury
TL;DR
CT liver-tumour segmentation matters for clinical assessment, but manual segmentation is time-consuming and routine size estimates rely on diameters despite volume better predicting survival. The paper uses two tandem FCNs trained end-to-end for joint liver and lesion segmentation. On LiTS, it ranked fourth in lesion segmentation and achieved competitive liver and lesion detection and segmentation results.
Problem
Manual CT tumour segmentation is time-consuming, and clinical practice usually estimates size from axial diameters even though tumour volume better predicts patient survival.
Source: §1 · p. 1Method
Two tandem, jointly trained FCNs segment liver and lesions, with the liver FCN's representation passed to the lesion FCN.
Source: §2.1 · p. 1, §3 · p. 3Results
The method ranked fourth in LiTS lesion segmentation with Dice per case 0.661; its liver Dice per case was 0.951, and its >50% overlap precision was 0.446 versus leHealth's 0.156.
Source: §3 · p. 3Takeaways & Limitations
The authors present the model's simplicity as a base for architectural research on liver and liver-lesion segmentation.
Source: §4 · p. 4Memory constraints limit the proposed model to processing 2D slices.
Source: §2.3 · p. 2
Abstract
from arXiv · showhide
We propose a model for the joint segmentation of the liver and liver lesions in computed tomography (CT) volumes. We build the model from two fully convolutional networks, connected in tandem and trained together end-to-end. We evaluate our approach on the 2017 MICCAI Liver Tumour Segmentation Challenge, attaining competitive liver and liver lesion detection and segmentation scores across a wide range of metrics. Unlike other top performing methods, our model output post-processing is trivial, we do not use data external to the challenge, and we propose a simple single-stage model that is trained end-to-end. However, our method nearly matches the top lesion segmentation performance and achieves the second highest precision for lesion detection while maintaining high recall.
1. INTRODUCTION
CT liver-tumour segmentation can support assessment and care, but manual segmentation is time-consuming and clinical size estimates usually use axial diameters even though volume better predicts survival. The paper presents a jointly trained, end-to-end model and reports competitive challenge performance.
1. INTRODUCTION
- Tumour volume better predicts patient survival than diameter, while manual CT segmentation is time-consuming and clinical practice usually estimates tumour size from axial-plane diameters.Segmentation is needed for tumour-load assessment, treatment planning, prognosis, and monitoring treatment response.Source: §1 · p. 1
- The authors build on fully convolutional networks to configure a model for segmenting metastatic liver lesions in CT volumes.They draw on an architecture evaluated in; related medical-image segmentation models are often UNet-like.Source: §1 · p. 1
- The model attains competitive liver and lesion detection and segmentation scores on the 2017 MICCAI Liver Tumour Segmentation Challenge, with trivial output post-processing and end-to-end single-stage training.The authors report using no data external to the challenge.Source: §1 · p. 1
2. METHOD
The method uses two tandem FCNs to segment liver and lesions, trains and evaluates on the LiTS dataset, and combines slice-level predictions with limited post-processing.
2.1. Model
- Two UNet-like FCNs are trained end-to-end: the first segments liver, and its latent representation is passed to the lesion-segmentation FCN.The FCN architecture uses contracting and expanding paths with skip connections; the expanding path adds corresponding contracting-path features to recover spatial detail.Source: §2.1 · p. 1, Figure 1 · p. 2
2.2. Data Set
- The dataset comprises 200 contrast-enhanced portal-venous CT volumes; 130 publicly released cases have manual liver and lesion segmentations, while 70 are held out for evaluation.Of the 130 annotated cases, 115 were used for training and 15 for validation; preprocessing was limited to intensity scaling and clipping.Source: §2.2 · p. 2
2.3. Training the Model
- Training used 2D axial slices containing liver, with augmentation and a two-stage resolution schedule: downscaled training followed by full-resolution fine-tuning.Three consecutive slices also provide cross-slice context: their pre-classifier outputs are combined by a 3x3 convolution to train a classifier for the middle slice.Source: §2.3 · p. 2
2.4. Generating Segmentations
- At test time, predictions are averaged across four flipped orientations and three similar models; the largest liver component is retained and lesion predictions are cropped to a liver mask dilated by 20 mm.No other output post-processing is performed.Source: §2.4 · p. 3
3. RESULTS AND DISCUSSION
On LiTS, the method ranked fourth in lesion segmentation and performed well on liver segmentation, while its detection precision compared favorably with some higher-ranked methods. The authors attribute the approach's distinction to end-to-end, single-stage conditioning of lesion segmentation on liver-network features.
3. RESULTS AND DISCUSSION
- The method ranked fourth in LiTS lesion segmentation with Dice per case 0.661, and liver segmentation achieved Dice per case 0.951 against the best entry's 0.963.Three methods attained higher mean lesion Dice per case; challenge entries were ranked by Dice per case.Source: §3 · p. 3
- At >50% overlap, leHealth's Dice per case 0.702 came with precision 0.156 versus this method's 0.446; hchen scored 0.686 with precision 0.409.The authors contrast their single-stage, challenge-data-only method and trivial post-processing with multi-stage, ensemble, post-processed, or externally trained alternatives.Source: §3 · p. 3
- Unlike other top methods that condition lesion segmentation on prior liver segmentation, this model passes the liver FCN's high-dimensional pre-classifier representation to the lesion FCN.The authors say this lets the lesion network focus on the liver and ignore distant lesions; they report that a single FCN for both tasks was less effective.Source: §3 · pp. 3–4
4. CONCLUSION
The authors present their end-to-end joint liver and lesion segmentation model as a simple base for further architectural research.
4. CONCLUSION
- The model performs end-to-end joint liver and lesion segmentation in CT without input-image preprocessing or complicated output post-processing, and its simplicity makes it a base for architectural research.The authors identify processing whole CT volumes rather than 2D slices as a possible route to improved segmentation performance.Source: §4 · p. 4