PaperlayerView exact revision on arXiv

Source-linked AI summary

Segmentation of Liver Lesions with Reduced Complexity Deep Models

Ram Krishna Pandey, Aswin Vasan, A G Ramakrishnan

arXiv:1805.09233v1cs.CV

TL;DR

Manual segmentation of each CT slice requires substantial human effort and time and is prone to errors, with results varying between oncologists. The paper extends UNet with residual connections, depthwise separable convolutions, and bilinear interpolation plus subpixel convolution for upsampling. The model achieves competitive LiTS challenge scores while using about 2 million learnable parameters instead of roughly 30 million for baseline UNet, a 13.8x reduction.

  • Problem

    Manual segmentation of each CT slice requires substantial human effort and time and is prone to errors, with results varying between oncologists.

    Source: §I · p. 1

  • Method

    The paper extends UNet with residual connections, depthwise separable convolutions, and bilinear interpolation plus subpixel convolution for upsampling.

    Source: §IV · p. 2, A. Architecture · p. 2

  • Results

    The model achieves competitive LiTS challenge scores while using about 2 million learnable parameters instead of roughly 30 million for baseline UNet, a 13.8x reduction.

    Source: §VI · p. 4

  • Takeaways & Limitations

    The architecture offers a compact, end-to-end trainable basis for faster inference and potential mobile medical-diagnosis applications.

    Source: §VI · p. 4, §VII · p. 4

Abstract

from arXiv · show

We propose a computationally efficient architecture that learns to segment lesions from CT images of the liver. The proposed architecture uses bilinear interpolation with sub-pixel convolution at the last layer to upscale the course feature in bottle neck architecture. Since bilinear interpolation and sub-pixel convolution do not have any learnable parameter, our overall model is faster and occupies less memory footprint than the traditional U-net. We evaluate our proposed architecture on the highly competitive dataset of 2017 Liver Tumor Segmentation (LiTS) Challenge. Our method achieves competitive results while reducing the number of learnable parameters roughly by a factor of 13.8 compared to the original UNet model.

I. INTRODUCTION

Liver-lesion segmentation from CT scans is laborious, variable across clinicians, and difficult for conventional methods, motivating efficient automated models. The paper targets a lightweight segmentation framework suitable for constrained devices and broader medical access.

I. INTRODUCTION

  • Manual segmentation of volumetric CT scans is time-consuming and error-prone, with results varying between oncologists.A scan can contain up to a thousand 2D slices, making manual processing especially demanding.Source: §I · p. 1
  • Medical lesion segmentation is challenging because CT appearance varies with shape, area, noise, position, and magnitude.Traditional thresholding, intensity clustering, and region-growing methods have shown poor performance on these images.Source: §I · p. 1
  • Encoder-decoder segmentation architectures extract coarse features before restoring input resolution through upsampling.State-of-the-art architectures often contain millions of parameters and require substantial compute power.Source: §I · p. 1
  • UNet-based encoder-decoder models are powerful baselines for medical image segmentation across CT and MRI tasks.Fully convolutional networks, including 2D and 3D variants, commonly serve as backbones for volumetric segmentation.Source: §I · p. 1
  • The paper develops a fast, efficient segmentation framework intended to run on mobile phones and assist diagnosis where computing resources and doctors are limited.The authors position lightweight models as tools to support doctors and scientific diagnosis.Source: §I · p. 1

II. RELATED WORK

Prior work largely builds on encoder-decoder and UNet-style architectures, but increasingly complex three-dimensional or multi-network designs can demand substantial computation. The paper instead emphasizes a simple, efficient single-model architecture optimized for mobile use.

II. RELATED WORK

  • Dense-prediction models commonly use encoder-decoder frameworks that preserve semantic and spatial information through specialized decoder designs.Decoder variants include skip concatenation, pooling-index reuse, and full encoder-feature concatenation.Source: §II · pp. 1–2
  • UNet variants have become widely used for medical-image segmentation, including lesion, organ, and LiTS challenge solutions [2, 12, 19, 23, 20].Reported alternatives include two-step UNets, ResNet blocks, 3D convolutions, hybrid 3D/2D models, and parallel fully convolutional networks.Source: §II · p. 2
  • Three-dimensional UNet operations can be GPU-intensive and require substantial processing power without substantial performance gains.The related work notes that leading LiTS contenders used two-dimensional UNet variants.Source: §II · p. 2
  • Depthwise separable convolutions reduce model footprint and have been used in mobile-oriented architectures with fewer parameters than state-of-the-art models [14].Their reduced footprint makes them suitable for mobile vision applications.Source: §II · p. 2
  • The proposed approach combines components from recent deep-learning developments into a simple single-model architecture that runs faster than traditional architectures and is optimized for mobile devices.This positioning connects the paper's design to the computational constraints identified in prior work.Source: §II · p. 2

A. Dataset

The study uses 130 contrast-enhanced volumetric CT scans from the LiTS training dataset, standardizes slice size, and applies CT-specific preprocessing to improve lesion visibility. The scans vary substantially in resolution, field of view, and depth.

A. Dataset

  • The LiTS training dataset contains 130 3D volumetric CT scans collected from different clinical sites.The scans are contrast-enhanced, and each scan contains roughly forty to more than a thousand slices.Source: A. Dataset · p. 2
  • Input slices are resized to 256 pixels per side despite variation in scanner resolution and field of view.The number of slices varies across scans, from around forty to over a thousand.Source: A. Dataset · p. 2

B. Preprocessing

  • CT windowing maps Hounsfield units to the range -100 to 200 HU to provide higher soft-tissue contrast for tumor-nodule detection.The operation highlights structures relevant to identifying lesions.Source: B. Preprocessing · p. 2
  • Histogram equalization enhances contrast by redistributing intensity frequencies and spacing closely placed values.It is used as an image-processing step before segmentation.Source: B. Preprocessing · p. 2

IV. CONTRIBUTION

The method extends UNet with residual connections, depthwise separable convolutions, and parameter-free bilinear-plus-subpixel upsampling. These changes target efficient training and inference while preserving segmentation performance.

IV. CONTRIBUTION

  • The proposed method extends UNet with recent deep-learning components as its main architectural contribution.The design combines multiple modifications rather than introducing a separate segmentation paradigm.Source: §IV · p. 2
  • Short-range residual connections improve gradient flow, alleviate vanishing gradients, and support smoother, more efficient training.The contribution specifically uses residual connections to address optimization difficulty.Source: §IV · p. 2
  • Depthwise separable convolutions reduce parameter count by a factor of over thirteen relative to traditional UNet without compromising performance.They replace standard convolutional operations as a central efficiency mechanism.Source: §IV · p. 2
  • Combining bilinear interpolation with subpixel convolution upsamples decoder features to the output-mask resolution, avoiding slow deconvolution layers while reducing learning parameters and increasing training and inference speed.Source: §IV · p. 2

V. METHOD

The method uses a lightweight encoder-decoder architecture combining depthwise separable convolutions, ResNet-like blocks, skip connections, and parameter-free upsampling.

A. Architecture

  • The model uses an encoder-decoder design based on UNet, replacing standard convolutions with depthwise separable convolutions to reduce its parameter footprint.The encoder contains four ResNet-like blocks with pooling, while the innermost representation reaches depth 1024.Source: A. Architecture · pp. 2–3
  • Bilinear interpolation performs the first three decoder upsampling steps, followed by subpixel convolution for the final resolution increase.The subpixel layer rearranges pixel values while increasing spatial resolution and reducing feature depth.Source: A. Architecture · p. 3
  • Short-range residual connections support gradient flow, while long-range copy-and-concatenate connections restore spatial information lost during pooling.Encoder feature maps are concatenated with corresponding upsampled decoder maps to improve mask reconstruction.Source: A. Architecture · p. 3, Figure 1 · p. 3
  • Table I presents the model's performance comparison with submissions to the LiTS Challenge.The supplied table caption identifies the comparison scope but does not provide cell values.Source: Table I · p. 3

B. Training and Implementation

The model was trained with Adam, augmentation, normalization, dropout, and validation over 100,000 iterations, with training and validation behavior tracked by Dice score and loss.

B. Training and Implementation

  • Training used Adam with learning rate 0.001, weighted cross-entropy loss, augmentation, batch normalization, dropout, and a 75:25 k-fold train-validation split.The model was trained for 100,000 iterations with batch size 16, and the best validation weights were retained.Source: B. Training and Implementation · p. 3
  • Training required approximately two days on a Titan X GPU, compared with around four days for the baseline UNet.The reported hardware had 12 GB of RAM.Source: B. Training and Implementation · p. 3
  • Figure 2 plots Dice score and loss for training and validation as functions of iteration count.The curves summarize behavior after training for 100,000 iterations.Source: Figure 2 · p. 4
  • Figure 3 compares input CT images, ground-truth masks, and segmentation outputs from the proposed model.The three columns are arranged as inputs, ground truths, and model results.Source: Figure 3 · p. 4

VI. RESULTS AND DISCUSSION

The proposed method achieves competitive LiTS segmentation performance while using substantially fewer learnable parameters than the baseline UNet.

VI. RESULTS AND DISCUSSION

  • Dice score measures overlap between predicted and ground-truth regions using true positives and false positives.The paper describes the score as the ratio involving intersection and union of the predicted and ground-truth sets.Source: §VI · p. 4
  • Although Vorontsov et. al. [20] achieve a higher Dice score, the proposed method compares favorably in parameter efficiency.The model also avoids pre-trained models and post-processing such as 3D Conditional Random Fields [18].Source: §VI · p. 4
  • 2 million learnable parameters achieve competitive performance, compared with around 30 million for baseline UNet, yielding a 13.8x reduction.The comparison is reported for the proposed model and the baseline UNet.Source: §VI · p. 4

VII. CONCLUSION

The architecture reduces model footprint and computational cost through lightweight operations while retaining an end-to-end trainable segmentation pipeline.

VII. CONCLUSION

  • Replacing deconvolution with bilinear interpolation and subpixel convolution reduces learnable parameters and speeds training and inference.Depthwise separable convolutions further reduce parameters, while skip connections support training and mask accuracy.Source: §VII · p. 4

VIII. FUTURE WORK

The authors identify decoder upsampling as an area for further exploration, including alternative interpolation, subpixel, and deconvolutional combinations. They also propose extending and optimizing the model for faster native mobile deployment.

  • The authors propose exploring alternative decoder upsampling techniques and combinations of interpolation, subpixel, and deconvolutional methods.Source: §VIII · p. 4
  • The model is presented as suitable for future mobile deployment because of its reduced memory footprint and fast inference time.Source: §VIII · p. 5
  • Further optimization is proposed to make the model perform faster operations.Source: §VIII · p. 5
Loading 1805.09233v1…