Source-linked AI summary
Automatic Liver Lesion Detection using Cascaded Deep Residual Networks
Lei Bi, Jinman Kim, Ashnil Kumar, Dagan Feng
TL;DR
Manual and traditional liver-lesion segmentation is subjective, time-consuming, and dependent on handcrafted features and user knowledge. The method uses deep residual networks and a cascaded ResNet with multi-scale fusion to progressively refine liver and lesion boundaries. The proposed method achieved an overall Dice index of 64.00% for liver lesion segmentation.
Problem
Manual and traditional liver-lesion segmentation is subjective, time-consuming, and dependent on handcrafted features and user knowledge.
Source: §1 · p. 1Method
The method uses deep residual networks and a cascaded ResNet with multi-scale fusion to progressively refine liver and lesion boundaries.
Source: §1 · p. 2, §2.4 · p. 3Results
The proposed method achieved an overall Dice index of 64.00% for liver lesion segmentation.
Source: §4 · p. 5Takeaways & Limitations
Deep residual and cascaded architectures improved segmentation over traditional VGGNet-based FCNs, including a 20.13% liver-lesion Dice improvement.
Source: §1 · p. 2, §3.2 · p. 4Traditional approaches remain unreliable because they depend on handcrafted features, user priors, low-level texture, and extensive parameter tuning.
Source: §1 · p. 1
Abstract
from arXiv · showhide
Automatic segmentation of liver lesions is a fundamental requirement towards the creation of computer aided diagnosis (CAD) and decision support systems (CDS). Traditional segmentation approaches depend heavily upon hand-crafted features and a priori knowledge of the user. As such, these methods are difficult to adopt within a clinical environment. Recently, deep learning methods based on fully convolutional networks (FCNs) have been successful in many segmentation problems primarily because they leverage a large labelled dataset to hierarchically learn the features that best correspond to the shallow visual appearance as well as the deep semantics of the areas to be segmented. However, FCNs based on a 16 layer VGGNet architecture have limited capacity to add additional layers. Therefore, it is challenging to learn more discriminative features among different classes for FCNs. In this study, we overcome these limitations using deep residual networks (ResNet) to segment liver lesions. ResNet contain skip connections between convolutional layers, which solved the problem of the training degradation of training accuracy in very deep networks and thereby enables the use of additional layers for learning more discriminative features. In addition, we achieve more precise boundary definitions through a novel cascaded ResNet architecture with multi-scale fusion to gradually learn and infer the boundaries of both the liver and the liver lesions. Our proposed method achieved 4th place in the ISBI 2017 Liver Tumor Segmentation Challenge by the submission deadline.
1. INTRODUCTION
The paper targets automated liver and liver-lesion segmentation because manual and traditional approaches are subjective, time-consuming, and dependent on handcrafted features. It proposes deep residual and cascaded ResNet architectures to learn discriminative features and iteratively refine boundaries.
1. INTRODUCTION
- Manual liver-lesion segmentation is subjective, poorly reproducible, and time-consuming, motivating fully automated methods.Source: §1 · p. 1
- Traditional methods rely on handcrafted features, user priors, low-level texture, and extensive parameter tuning, making accurate segmentation unreliable.Source: §1 · p. 1
- FCNs learn low-level appearance and high-level semantic features end-to-end, but VGGNet-based FCNs are limited in depth and produce coarse, spatially inconsistent lesion boundaries.Source: §1 · p. 1
- ResNet uses shortcut connections to reduce training degradation in very deep networks and support more discriminative feature learning.Source: §1 · pp. 1–2
- The study introduces an automatic liver and liver-tumor detection method based on deep residual networks that improves liver and lesion detection accuracy over VGGNet-based FCNs.Source: §1 · p. 2
- A cascaded ResNet iteratively uses previous estimates during training and testing to refine liver and lesion boundaries.Source: §1 · p. 2
2. METHODS
The study evaluates a cascaded ResNet pipeline for liver and lesion segmentation using LiTS CT studies, residual learning, probability-map refinement, and multi-scale inference.
2.1. Materials
- LiTS provides 201 contrast-enhanced CT studies from six medical centers, split into 118 training and 13 validation studies from the 131-study training set.The dataset also includes 70 test images and expert-supervised liver and lesion masks for training studies.Source: §2.1 · p. 2
2.2. Pre-processing
- Voxel intensities are restricted to [-160, 240] HU and normalized to [0, 1] to exclude irrelevant structures while retaining liver lesions.The range extends below the cited liver window to help capture lesions.Source: §2.2 · p. 2
2.3. Deep Residual Networks for Segmentation
- Residual blocks use shortcut connections around convolutional layers, enabling deeper networks without training degradation and providing more discriminative features.The segmentation model adds convolutional and deconvolutional layers for upsampling and dilates ResNet feature maps to create score masks.Source: §2.3 · p. 2
2.4. Cascaded Deep Residual Networks for Segmentation
- The cascaded ResNet embeds probability maps produced by previous deep residual networks during training and testing to predict tumor and liver regions.During testing, outputs from image sizes 512×512 to 640×640 in 32-pixel increments were averaged; morphological filtering filled holes, with no other post-processing.Source: §2.4 · pp. 2–3
2.5. Implementation Details
- Training uses 8,802 selected 2D axial slices, ImageNet fine-tuning, stochastic gradient descent, augmentation, and two Titan X GPUs for approximately seven days.Testing uses a ResNet stage to produce liver and tumor probability maps before cascaded prediction; inference takes approximately 48 minutes multi-scale or 16 minutes single-scale per average CT volume.Source: §2.5 · pp. 3–4
3. EXPERIMENTS AND RESULTS
The validation study compares VGGNet FCN, ResNet, cascaded ResNet, 3D-CRF post-processing, and multi-scale fusion on liver and lesion segmentation. The cascaded ResNet with multi-scale fusion achieved the best segmentation results, improving Dice accuracy over VGGNet FCN for both tasks.
3.2. Results
- The cascaded ResNet with multi-scale fusion achieved the best segmentation results among the evaluated methods.Evaluation used Dice and Jaccard indices on 13 validation studies because test ground truth was unavailable when the manuscript was prepared.Source: §3.2 · p. 4, §3.1 · p. 4
- 3.94% higher Dice for liver segmentation and 20.13% higher Dice for liver lesion segmentation were achieved versus traditional VGGNet-based FCN.Source: §3.2 · p. 4
4. DISCUSSIONS
The cascaded ResNet with multi-scale fusion achieved the best reported segmentation results and ranked fourth in the challenge submission, with a 64.00% overall lesion Dice index.
- The cascaded ResNet with multi-scale fusion achieved the best results among the compared methods.Source: §4 · p. 4, Table 2 · p. 5
- Cascading improved liver lesion segmentation by 1.33% in Dice index over the base configuration.Source: §4 · p. 4
- The cascaded architecture iteratively refines liver and lesion boundaries using high-level semantic differences rather than shared low-level pixel values.Source: §4 · p. 4
- Multi-scale fusion was attributed to the strongest overall performance because it is scale-invariant across studies with differing staging and pixel resolutions.Source: §4 · p. 4
- 64.00% overall Dice index for liver lesion segmentation accompanied the method’s fourth-place online leaderboard ranking by the submission deadline.Source: §4 · p. 5