Source-linked AI summary
A Novel Focal Tversky loss function with improved Attention U-Net for lesion segmentation
Nabila Abraham, Naimul Mefraz Khan
TL;DR
Medical lesion segmentation is challenged by severe class imbalance and small regions, motivating methods that better balance precision and recall. The paper proposes a focal Tversky loss with an attention U-Net enhanced by an image pyramid, and reports higher Dice scores than baseline U-Net on BUS and ISIC 2018. The method is evaluated under a specific loss weighting and model-training setup.
Problem
Small lesion regions and class imbalance make it difficult to balance precision and recall with standard segmentation losses.
Method
The paper combines a focal Tversky loss with an attention U-Net incorporating a multi-scaled input image pyramid.
Results
25.7% and 3.6% Dice-score improvements over baseline U-Net were reported on BUS dataset B and ISIC 2018, respectively.
Takeaways & Limitations
The proposed method outperforms baseline U-Net in Dice scores while producing more balanced precision-recall scores with low standard deviations.
Takeaways & Limitations
The reported training configuration fixes α = 0.7 and β = 0.3, while the loss behavior depends on these weighting choices.
Abstract
from arXiv · showhide
We propose a generalized focal loss function based on the Tversky index to address the issue of data imbalance in medical image segmentation. Compared to the commonly used Dice loss, our loss function achieves a better trade off between precision and recall when training on small structures such as lesions. To evaluate our loss function, we improve the attention U-Net model by incorporating an image pyramid to preserve contextual features. We experiment on the BUS 2017 dataset and ISIC 2018 dataset where lesions occupy 4.84% and 21.4% of the images area and improve segmentation accuracy when compared to the standard U-Net by 25.7% and 3.6%, respectively.
1. INTRODUCTION
Medical lesion segmentation must handle severe class imbalance and small regions of interest. The paper addresses this with a focal Tversky loss and an improved attention U-Net, achieving higher Dice scores than standard U-Net baselines on two datasets.
- Small pathological regions occupy only a minor fraction of medical images, creating a central segmentation challenge.
- Focal loss reduces the influence of numerous easy negative examples to alleviate class imbalance.
- Attention gates focus on target regions and suppress activations from irrelevant regions during end-to-end training.
- The proposed focal Tversky loss targets precision-recall balance for small ROIs, while a multi-scaled input pyramid improves intermediate feature representations.
- 25.7% and 3.6% Dice-score improvements over baseline U-Net were reported on BUS dataset B and ISIC 2018, respectively.
2. METHODOLOGY
The methodology combines a focal Tversky loss, which emphasizes hard and false-negative lesion errors, with an attention U-Net enhanced by multiscale inputs and deep supervision.
- Focal Tversky Loss: The Dice loss weights false positives and false negatives equally, whereas the Tversky index allows their contributions to be balanced for small lesions.The method emphasizes false negatives more heavily than false positives to improve recall under severe class imbalance.
- Focal Tversky Loss: The focal Tversky loss applies an exponent γ to focus training on hard, low-Tversky examples while reducing the contribution of easy examples.The exponent is described as varying from 1 to 3 in the formulation, while experiments report best performance with γ = 4.
- Focal Tversky Loss: The experiments use α = 0.7 and β = 0.3, while α = β = 0.5 reduces the Tversky index to the Dice score and γ = 1 reduces FTL to TL.A higher α is hypothesized to shift emphasis toward minimizing false-negative predictions.
- Network Architecture: The improved Attention U-Net combines contracting and expansive paths, skip connections, soft attention gates, and the focal Tversky loss.Attention gates use coarse contextual signals to identify relevant spatial information in low-level feature maps before decoding.
- Network Architecture: An input image pyramid is injected before max-pooling layers, and deep supervision preserves multiscale and locality-aware representations for small ROI segmentation.Intermediate feature representations are supervised with FTL, except for the last layer, which uses Tversky loss to avoid over-suppression.
3. EXPERIMENTS
Experiments evaluate focal Tversky loss and multi-scaled attention U-Net variants on BUS 2017 and ISIC 2018, using controlled comparisons against a Dice-loss U-Net baseline.
- BUS 2017 contains 163 breast-lesion ultrasound images, resampled to 128 x 128 pixels with a 75-25 train-test split.
- ISIC 2018 supplies training data from a larger skin-lesion dataset for extending the proposed method beyond BUS 2017.
- The evaluation compares seven U-Net and Tversky-loss variations with a baseline U-Net trained using Dice loss.
- Ablation results use 5fold cross validation and report Dice scores, precision, and recall.
- Training uses 50 epochs for ISIC 2018, 100 for BUS 2017, stochastic gradient descent with momentum, and grid-searched parameters.
4. RESULTS
The proposed combinations of focal Tversky loss and improved attention U-Net outperform the baseline in the reported experiments, with effects differing between BUS and ISIC datasets.
- BUS 2017: 0.804 Dice score is achieved by the proposed architecture when trained on a small subset of 100 BUS images.
- BUS 2017: The baseline U-Net with Dice loss has the worst BUS performance and unstable precision and recall, while focal Tversky variants improve DSC balance.
- BUS 2017: Input-pyramid integration significantly improves attention U-Net DSC, suggesting that small-lesion features are otherwise easily lost under high class imbalance.
- ISIC 2018: 3.6% improvement over the baseline is obtained on ISIC 2018, with a low spread of 0.7%.
- ISIC 2018: The improved attention U-Net with focal Tversky loss achieves the best ISIC DSC through better-balanced recall and precision.
5. CONCLUSION
The paper concludes that focal Tversky loss and an input image pyramid improve lesion segmentation by balancing precision and recall and preserving contextual information for small regions of interest.
- The focal Tversky loss is proposed to improve precision-recall balance in semantic segmentation.
- Experiments demonstrate that loss-function choice matters for highly imbalanced problems and datasets of varying sizes.
- An input image pyramid at each model scale adds redundant features that help recover contextual information lost during small-ROI segmentation.
- The proposed method outperforms the baseline U-Net in Dice scores and provides balanced precision-recall scores with low standard deviations.