A Novel Focal Tversky loss function with improved Attention U-Net for lesion segmentation

18 Oct 2018  Â·  Nabila Abraham, Naimul Mefraz Khan ·

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.

PDF Abstract

Datasets


  Add Datasets introduced or used in this paper
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Lesion Segmentation BUS 2017 Dataset B Attn U-Net + Multi-Input + FTL Dice Score 0.804 # 1
Lesion Segmentation BUS 2017 Dataset B Attn U-Net + DL Dice Score 0.615 # 4
Lesion Segmentation BUS 2017 Dataset B U-Net + FTL Dice Score 0.669 # 3
Lesion Segmentation ISIC 2018 Attn U-Net + DL Dice Score 0.806 # 14
Lesion Segmentation ISIC 2018 Attn U-Net + Multi-Input + FTL Dice Score 0.856 # 10
Lesion Segmentation ISIC 2018 U-Net + FTL Dice Score 0.829 # 13

Methods