RFS+: A Clinically Adaptable and Computationally Efficient Strategy for Enhanced Brain Tumor Segmentation

Abstract

Automated brain tumor segmentation has significant importance, especially for disease diagnosis and treatment planning. The study utilizes a range of MRI modalities, namely T1-weighted (T1), T1-contrast-enhanced (T1ce), T2-weighted (T2), and fluid-attenuated inversion recovery (FLAIR), with each providing unique and vital information for accurate tumor localization. While state-of-the-art models perform well on standardized datasets like the BraTS dataset, their suitability in diverse clinical settings (matrix size, slice thickness, manufacturer-related differences such as repetition time, and echo time) remains a subject of debate. This research aims to address this gap by introducing a novel ‘Region-Focused Selection Plus (RFS+)’ strategy designed to efficiently improve the generalization and quantification capabilities of deep learning (DL) models for automatic brain tumor segmentation. RFS+ advocates a targeted approach, focusing on one region at a time. It presents a holistic strategy that maximizes the benefits of various segmentation methods by customizing input masks, activation functions, loss functions, and normalization techniques. Upon identifying the top three models for each specific region in the training dataset, RFS+ employs a weighted ensemble learning technique to mitigate the limitations inherent in each segmentation approach. In this study, we explore three distinct approaches, namely, multi-class, multi-label, and binary class for brain tumor segmentation, coupled with various normalization techniques applied to individual sub-regions. The combination of different approaches with diverse normalization techniques is also investigated. A comparative analysis is conducted among three U-net model variants, including the state-of-the-art models that emerged victorious in the BraTS 2020 and 2021 challenges. These models are evaluated using the dice similarity coefficient (DSC) score on the 2021 BraTS validation dataset. The 2D U-net model yielded DSC scores of 77.45%, 82.14%, and 90.82% for enhancing tumor (ET), tumor core (TC), and the whole tumor (WT), respectively. Furthermore, on our local dataset, the 2D U-net model augmented with the RFS+ strategy demonstrates superior performance compared to the state-of-the-art model, achieving the highest DSC score of 79.22% for gross tumor volume (GTV). The model utilizing RFS+ requires 10% less training dataset, 67% less memory and completes training in 92% less time compared to the state-of-the-art model. These results confirm the effectiveness of the RFS+ strategy for enhancing the generalizability of DL models in brain tumor segmentation.