Source-linked AI summary
HarDNet-MSEG: A Simple Encoder-Decoder Polyp Segmentation Neural Network that Achieves over 0.9 Mean Dice and 86 FPS
Chien-Hsiang Huang, Hung-Yu Wu, Youn-Long Lin
TL;DR
Polyp segmentation supports colonoscopy-based colorectal cancer prevention, but manual and existing automated methods face efficiency and accuracy challenges. HarDNet-MSEG combines a HarDNet68 backbone with a partial-decoder encoder-decoder design. It achieves state-of-the-art accuracy and inference speed across five datasets, including 0.904 mean Dice at 86.7 FPS on Kvasir-SEG.
Problem
Polyp detection and segmentation are difficult because manual detection is time-consuming, earlier methods miss polyps, and CNN methods can have boundary, small-region, fragmentation, and speed problems.
Method
HarDNet-MSEG uses a simple encoder-decoder architecture with HarDNet68 as the backbone and a decoder inspired by the Cascaded Partial Decoder.
Results
HarDNet-MSEG achieves state-of-the-art performance across five challenging datasets, exceeding 0.90 mean Dice on Kvasir-SEG and running 1.3 times faster than PraNet.
Takeaways & Limitations
The results support HarDNet as an efficient backbone for medical imaging segmentation and show that a simple encoder-decoder can deliver high accuracy with efficient inference.
Abstract
from arXiv · showhide
We propose a new convolution neural network called HarDNet-MSEG for polyp segmentation. It achieves SOTA in both accuracy and inference speed on five popular datasets. For Kvasir-SEG, HarDNet-MSEG delivers 0.904 mean Dice running at 86.7 FPS on a GeForce RTX 2080 Ti GPU. It consists of a backbone and a decoder. The backbone is a low memory traffic CNN called HarDNet68, which has been successfully applied to various CV tasks including image classification, object detection, multi-object tracking and semantic segmentation, etc. The decoder part is inspired by the Cascaded Partial Decoder, known for fast and accurate salient object detection. We have evaluated HarDNet-MSEG using those five popular datasets. The code and all experiment details are available at Github. https://github.com/james128333/HarDNet-MSEG
1 Introduction
Colorectal cancer prevention depends on regular colonoscopy and polyp removal, but manual and earlier automated segmentation remain difficult, error-prone, or slow. HarDNet-MSEG addresses these challenges with a simple encoder-decoder architecture based on HarDNet68.
- 95% of CRC is due to a colorectal adenomatous polyp, making regular colonoscopy and early treatment important.
- Manual polyp detection is time-consuming and depends heavily on the endoscopist’s experience and ability.
- Earlier feature-and-classifier methods have high missed-detection rates because polyps vary in position, size, and color.
- CNN-based methods improve polyp segmentation, but still face boundary errors, missing small regions, fragmented large areas, and long inference times.
- HarDNet-MSEG uses a simple encoder-decoder architecture with a HarDNet68 backbone to combine high accuracy with efficient inference on polyp-segmentation benchmarks.
2 Related work
Medical image segmentation has progressed from FCN and U-Net architectures toward stronger backbones and specialized modules. HarDNet-MSEG combines HarDNet with an encoder-decoder design and reports high accuracy with efficient inference across five polyp datasets.
- FCN introduced end-to-end pixel classification, while U-Net established a widely used encoder-decoder baseline with multiscale skip connections for medical imaging.
- Stronger CNN backbones and modules such as spatial pyramid pooling and attention improve recognition, receptive fields, and multiscale information integration.
- HarDNet-MSEG uses HarDNet as its backbone in an encoder-decoder architecture and achieves high accuracy with efficient inference on five named polyp datasets.
- The overview identifies HarDNet68 as the encoder and a partial decoder as the decoder components of HarDNet-MSEG.
3 HarDNet-MSEG
HarDNet-MSEG uses HarDNet68 as a memory-efficient encoder and a partial decoder that prioritizes deeper features while aggregating multiscale information. Its design adds receptive-field expansion and feature aggregation without using HarDBlock in the decoder.
- 3.1 Backbone : HarDNet: HarDNet reduces memory traffic by shortening dense-block shortcuts, widening key-layer channels, and using limited 1x1 convolutions to increase computational density.
- 3.1 Backbone : HarDNet: 30% inference time reduction versus DenseNet and ResNet accompanies higher ImageNet accuracy for HarDNet’s backbone design.
- 3.1 Backbone : HarDNet: HarDNet68 is used as the backbone for colorectal-polyp semantic segmentation.
- 3.2 Cascaded Partial Decoder: The decoder omits HarDBlock to balance inference time and performance, distinguishing it from FC-HarDNet.
- 3.2 Cascaded Partial Decoder: The partial decoder discards shallow features, computes more on deeper features, and aggregates scales through convolutions and skip connections.
- 3.2.1 RFB Module: The RFB module uses multibranch standard and dilated convolutions, then a 1x1 convolution, to produce representations with different receptive fields.
- 3.2.1 RFB Module: RFB modules are added to skip connections to enlarge receptive fields across feature maps at different resolutions.
- 3.2.2 Dense Aggregation: Aggregation upsamples features to a common scale and combines corresponding feature maps through element-wise multiplication.
4 Experiments
The experiments compare HarDNet-MSEG with state-of-the-art models across five polyp-segmentation datasets under two training-data settings, emphasizing accuracy and inference speed. Across these evaluations, the model achieves strong metric performance and efficient inference on an RTX 2080 Ti GPU.
- Two experiments reproduce the distinct training-data splits and methods used in prior studies to reduce variable factors during comparison.The authors use training data from references [13] and [20] and follow their respective training methods.
- Five datasets—Kvasir-SEG, CVC-ColonDB, EndoScene, ETIS-Larib Polyp DB, and CVCClinic DB—are used for comparison with other state-of-the-art models.
- 880 Kvasir-SEG images are used for training and 120 for testing in one setting, with augmentation, 512x512 inputs, SGD, 100 epochs, and a 1e-2 learning rate.
- HarDNet-MSEG shows the greatest accuracy on most metrics in Table 1 and substantially faster inference than the compared models.
- 1450 training images are used without augmentation in the second setting, and HarDNet-MSEG achieves the best mean Dice and mIoU on each dataset with 88 FPS inference.The training images include 900 Kvasir-SEG and 550 CVC-ClinicDB images; inference is evaluated at 312x312 inputs.
- Mean Dice and Mean IoU are the primary comparison metrics, while additional reported metrics and platform details support the speed evaluation.Speed evaluations use an Intel i9-9900K CPU, GeForce RTX 2080 Ti, PyTorch 1.6, and CUDA 10.2.
5 Conculsion
HarDNet-MSEG reaches state-of-the-art performance across five challenging polyp-segmentation datasets while maintaining efficient inference. On Kvasir-SEG, it exceeds 0.90 mean Dice and produces improved boundary and prediction results in the reported examples.
- HarDNet-MSEG achieves state-of-the-art performance across all five challenging datasets.
- 0.912 mean Dice against [13] and 0.904 against [20] are reported on Kvasir-SEG, making HarDNet-MSEG the only network above 0.90 in that comparison.
- 1.3 times faster than PraNet and more than 2 times faster than other models are reported for HarDNet-MSEG.
- The model uses a simple encoder-decoder architecture without the attention modules used in [13] and.
- The Kvasir-SEG inference examples show better boundaries and more accurate predictions for HarDNet-MSEG.