Size and Smoothness Aware Adaptive Focal Loss for Small Tumor Segmentation

TL;DR

The paper addresses the challenge of segmenting small, irregular tumors in MRI by introducing Adaptive Focal Loss (A-FL), which dynamically modulates focusing strength and class balancing using tumor volume and boundary smoothness. The method integrates three adaptive factors—$\alpha_{va}$, $\gamma_{va}$, and $\gamma_{mSa}$—into the focal loss, yielding an adaptive modulating parameter $\gamma_{adaptive}$ and the loss $\text{A-FL}(P_t) = \alpha_{va} (1 - P_t)^{\gamma_{adaptive}} \log(P_t)$ with $\alpha_{va} = \dfrac{P_{bg}}{P_{fg} + P_{bg}}$, $\gamma_{va} = \dfrac{P_{fg}}{P_{fg} + P_{bg}}$, and $\gamma_{adaptive} = \gamma_{va} + \gamma_{mSa}$. evaluated on PICAI 2022 and BraTS 2018 using a ResNet50-based U-Net, A-FL consistently improves IoU and DSC over standard losses and other baselines, and shows pronounced gains for small and irregular tumors with better boundary delineation. While slightly more computationally expensive, the approach yields statistically robust improvements across multiple models and datasets, indicating strong potential for clinical 3D tumor segmentation tasks. The work also provides a detailed ablation and analysis of size- and boundary-related effects, supporting the practical value of adaptive loss tuning in medical image segmentation.

Abstract

Deep learning has achieved remarkable accuracy in medical image segmentation, particularly for larger structures with well-defined boundaries. However, its effectiveness can be challenged by factors such as irregular object shapes and edges, non-smooth surfaces, small target areas, etc. which complicate the ability of networks to grasp the intricate and diverse nature of anatomical regions. In response to these challenges, we propose an Adaptive Focal Loss (A-FL) that takes both object boundary smoothness and size into account, with the goal to improve segmentation performance in intricate anatomical regions. The proposed A-FL dynamically adjusts itself based on an object's surface smoothness, size, and the class balancing parameter based on the ratio of targeted area and background. We evaluated the performance of the A-FL on the PICAI 2022 and BraTS 2018 datasets. In the PICAI 2022 dataset, the A-FL achieved an Intersection over Union (IoU) score of 0.696 and a Dice Similarity Coefficient (DSC) of 0.769, outperforming the regular Focal Loss (FL) by 5.5% and 5.4% respectively. It also surpassed the best baseline by 2.0% and 1.2%. In the BraTS 2018 dataset, A-FL achieved an IoU score of 0.883 and a DSC score of 0.931. Our ablation experiments also show that the proposed A-FL surpasses conventional losses (this includes Dice Loss, Focal Loss, and their hybrid variants) by large margin in IoU, DSC, and other metrics. The code is available at https://github.com/rakibuliuict/AFL-CIBM.git.

Figures (5)

• Figure 1: Overall working pipeline comprises of three main parts: a. Data pre-processing; b. Segmentation Network; c. proposed Adaptive Loss Function (A schematic overview of the training process with our Adaptive Loss function (A-FL).
• Figure 2: The U-Net network architecture uses pre-trained ResNet50 he2016deep as backbone. The residual blocks in ResNet50 enable the network to learn more complex features and deeper representations, which are crucial for accurate segmentation.
• Figure 3: Performance comparison of regular Focal Loss (FL) vs. Adaptive Focal Loss (A-FL) on PICAI dataset. The left chart shows average Dice Similarity Coefficient (DSC) for large, medium, and small volume cases.The right chart shows average DSC for good, medium, and poor smoothness cases. Results highlight A-FL's effectiveness in improving segmentation accuracy, especially for small volume and irregularly shaped tumors.
• Figure 4: Performance comparison of regular Focal Loss (FL) vs. Adaptive Focal Loss (A-FL) on BraTs dataset. The left chart shows average Dice Similarity Coefficient (DSC) for large, medium, and small volume cases.The right chart shows average DSC for good, medium, and poor smoothness cases. Results highlight A-FL's effectiveness in improving segmentation accuracy, especially for small volume and irregularly shaped tumors.
• Figure 5: Qualitative results for the PICAI and BraTs validation sets are shown in 3D (a) and 2D slice views (b). The Label row shows ground truth, the A-FL row shows pblackictions from our method, and the FL row shows baseline Focal Loss pblackictions. For PICAI, (a) presents large (a1) and small (a2) volumes, along with good (a3) and poor (a4) smoothness. In (b), the $14^{th}$ and $2^{nd}$ slices show large (b1) and small (b2) volumes, while the $9^{th}$ and $7^{th}$ slices show good (b3) and poor (b4) smoothness. For BraTs, (a) similarly shows large and small volumes, with good and poor smoothness. In (b), the $111^{th}$ and $57^{th}$ slices depict large (b1) and small (b2) volumes, and the $50^{th}$ and $94^{th}$ slices show good (b3) and poor (b4) smoothness.