TSA-CA: Tumor-Scale Aware Curriculum Augmentation for Post-Treatment Glioma Segmentation

Abstract

Post-treatment glioma segmentation in brain MRI is both clinically significant and algorithmically challenging in computational neuro-oncology. Surgery, chemoradiotherapy, and adjuvant treatment substantially alter MRI characteristics. Resection cavities may mimic residual disease, radiation-induced necrosis can appear indistinguishable from recurrence, and small residual enhancing foci may occupy fewer than 100 voxels. Previous work [1] established a 3D U-Net baseline on BraTS 2024 post-treatment data using T1CE and FLAIR as independent input channels to segment Enhancing Tissue (ET), Surrounding Non-Enhancing T2/FLAIR Hyperintensity (SNFH), Non-Enhancing Tumor Core (NETC) and Resection Cavity (RC). While effective for large tumors, this framework systematically failed for small residual disease (total tumor volume <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$&lt;1,000$</tex> voxels) due to class imbalance, sparse training exposure, and feature-scale mismatch. We present Tumor-Scale Aware Curriculum Augmentation (TSA-CA), a five-stage cumulative training framework. (1) Modality-Robust Classical Augmentation applies per-channel spatial and intensity transforms. (2) Temperature-Annealed Volume Curriculum orders training by tumor-size difficulty via softmax scheduling, delivering the largest single gain (+0.063 overall Dice). (3) Scale-Adaptive Spatial Focus applies tumor-centric zoom and size-conditioned elastic deformation for small tumors, achieving the best boundary precision (HD95: <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{4. 0 2 ~ m m). }$</tex> Stages 1-3 raise ET DSC from <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.435 \pm 0.015$</tex> to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.550 \pm 0.020$</tex>, representing the core validated contribution of this work. (4) Adaptive Hard Example Distillation reveals an unexpected interaction with the Stage 3 training regime, analysed as a diagnostic contribution in the Discussion. (5) Generative Tumor Synthesis via TumorCut-and-Paste introduces 3D multi-modal copy-paste with per-modality intensity matching and Gaussian boundary blending; its independent evaluation is reserved as future work. We propose per-bin stratified reporting (Small/Medium/Large) as a standard supplement to aggregate metrics for BraTS 2024 post-treatment evaluation.