Source-linked AI summary
FocusNet: An attention-based Fully Convolutional Network for Medical Image Segmentation
Chaitanya Kaul, Suresh Manandhar, Nick Pears
TL;DR
Medical image segmentation needs effective deep architectures that can improve feature representations without extensive preprocessing. FocusNet adds a separate convolutional autoencoder-based attention branch to a convolutional network, achieving competitive results across skin cancer and lung segmentation benchmarks while improving on stated baselines.
Problem
Deep CNN accuracy can stagnate, while medical image segmentation often relies on preprocessing to achieve high accuracy.
Method
FocusNet uses a parallel encoder-decoder attention branch that learns hierarchical latent attention maps and gates feature representations in the segmentation network.
Results
FocusNet outperforms recurrent U-Net variants on lung segmentation except sensitivity and surpasses LIN in melanoma Jaccard Index without pre- or post-processing.
Takeaways & Limitations
The architecture provides competitive medical image segmentation performance with minimal preprocessing across varying image conditions.
Takeaways & Limitations
The network is less responsive to the sensitivity metric, which the authors partly attribute to training without pre- or post-processing.
Abstract
from arXiv · showhide
We propose a novel technique to incorporate attention within convolutional neural networks using feature maps generated by a separate convolutional autoencoder. Our attention architecture is well suited for incorporation with deep convolutional networks. We evaluate our model on benchmark segmentation datasets in skin cancer segmentation and lung lesion segmentation. Results show highly competitive performance when compared with U-Net and it's residual variant.
1. INTRODUCTION
The paper introduces an attention-based hybrid architecture that improves on a ResNet + SE model while requiring minimal preprocessing for medical image segmentation.
- CNNs provide end-to-end feature learning and have become baselines for computer vision, but accuracy can stagnate in very deep networks.
- FocusNet's attention mechanism improves on an already state-of-the-art ResNet + SE architecture.
- The architecture is designed for medical image processing with minimal preprocessing and images under varying conditions.The paper contrasts this with prior work using histogram equalization, denoising, or contrast equalization.
2. ATTENTION ARCHITECTURE
FocusNet uses parallel encoder-decoder branches, with one branch learning hierarchical latent attention maps that gate feature representations in the segmentation network.
- FocusNet uses two parallel encoder-decoder branches, with one branch devoted solely to attention and skip connections supporting gradient flow.The design biases the branches toward learning different representations.
- The attention branch hierarchically extracts latent attention maps from an encoder-decoder structure to support more accurate decoding.
- Each layer combines decoded attention outputs with feature maps through sigmoid gating and multiplication, as illustrated in the network architecture.
- The network uses residual convolutions, skip connections, strided-convolution downsampling, and SE blocks to support deeper training and recalibrate intermediate feature maps.
3. EVALUATION
The evaluation tests FocusNet on skin cancer and lung segmentation datasets using standard overlap, accuracy, sensitivity, and specificity metrics under a fixed training procedure.
- FocusNet is evaluated on skin cancer lesion segmentation and lung segmentation datasets.The skin cancer dataset includes 2000 training, 150 validation, and 600 test RGB images; the lung dataset contains 267 single-channel 512x512 images.
- Training uses Dice-coefficient loss, Adam optimization, plateau-based learning-rate reduction, validation monitoring, and a maximum of 80 epochs.
- Example melanoma results show good segmentation but indicate that preprocessing could improve accuracy.
- Data augmentation expands the lung training set with random zooms and flips and the skin cancer training set with channel shifts.The augmented sets contain 1700 lung images and 6000 skin cancer images.
- The evaluation reports Dice, Jaccard, accuracy, sensitivity, and specificity for all networks.Sensitivity is defined as the true positive rate and specificity as the true negative rate.
4. RESULTS
FocusNet outperforms recurrent U-Net variants on lung segmentation except sensitivity and is competitive on melanoma, surpassing LIN in Jaccard Index without pre- or post-processing.
- FocusNet outperforms R2U-Net and related recurrent architectures on every reported lung metric except sensitivity.Table 1 reports validation-set segmentation results for the lung dataset and extends prior results with FocusNet.
- On the melanoma test set, FocusNet is comparable with recent results and outperforms LIN in Jaccard Index without pre- or post-processing.The paper attributes this comparison to the second branch's bias toward learning more robust features, whereas LIN relied heavily on image preprocessing.
5. CONCLUSIONS
FocusNet uses attention-based encoding, residual convolutions, and skip connections to produce robust segmentation maps. Its sensitivity is weaker, which the authors partly attribute to training without pre- or post-processing.
- FocusNet combines attention-based encoding, residual convolutions, and skip connections to support robust segmentation and deeper-network training.The authors also state that skip connections facilitate gradient flow and residual convolutions help prevent overfitting.
- The network’s main limitation is reduced responsiveness to the sensitivity metric.The authors partly attribute this limitation to training without dataset pre- or post-processing.
- The attention mechanism is presented as generalizable beyond medical-image segmentation to other computer-vision domains.