Brain Tumor Segmentation via Noised Multi-level Wavelet Feature Extraction.

• POSTECH CSED499A - Research Project 1
• in MIV Lab @ POSTECH (Under the supervision by Wonhwa Kim, Hyuna Cho)
• (As of 12/15) All codes and docs are completed version.

• Implemented ‘WaveUNet’, that is able to extract useful features in image, with multi-level wavelet transform and various diffusion noise scales
• Proposed ‘Symmetric Contrastive Loss’, simple but strong logic
• Based on AttentionUNet, added a symmetric contrastive loss and had a better performance for tumor cores (NCR, ET).
• Feature extractor’s feature matching loss is hard to converge, need to modify the architecture or loss function in later

• Simple variation of wavelet transform, involving Discrete Wavelet Transform (DWT) & Discrete Inverse Wavelet Transform (IWT)
• $L = {1\over\sqrt(2)} [1 \ \ 1], H = {1\over\sqrt(2)} [-1 \ \ 1]$ represent low-pass & high-pass filters, construct 4 kernels ($LL^T, LH^T, HL^T, HH^T$)
• Decompose the input 𝑋 ∈ $R^{H \times W}$ into 4 subbands ($X_{ll}, X_{hl}, X_{hl}, X_{hh}$) with dimensions $R^{{H\over 2} \times {W\over 2}}$
• Accurate reconstruction of the original signals 𝑋 from frequency components through IWT

1. Input image is decomposed by Multi-level DWT, first low-frequency subbands 𝑋1,𝑙𝑙 gets noise and goes into WaveUNet
2. Higher level’s low-frequency subbands are concatenated residually into WaveUNet’s layers
3. WaveUNet is trained to mimics each level’s low frequency subbands, returns to image domain by IWT

• Can apply any general segmentation models (MLP, CNN, …)
• WaveUNet gives extracted multi-level feature map to Feature Voter, makes a per class score map

1. Feature matching Loss

• Feature extractor is learned to reduce the L2 norm of multilevel wavelet inputs & feature maps it extracts 2. Segmentation Loss
• Weighted multi-class cross-entropy loss
• Class distribution was highly imbalanced, weights as reciprocal of the class distribution for stable learning

• Tumor regions are almost asymmetric.
• If there is a tumor in one region based on the x-axis, there is very likely to no tumor in the opposite.
• Penalize not to be similar model’s prediction to opposite region’s label (Just tumor region)

• Used baseline model (AttentionUNet) learns to compare the effects by adding symmetric contrastive loss
• Each model trained for 100 epochs, hyperparameter 𝝀 = 0, 0.1, 0.3, 0.5
• Trained on 12,510 train datasets, and measure Dice Score for each classes with 400 validation images.

• Feature extraction with WaveUNet, give feature maps to AttentionUNet, λ=0.1 -> Not going well…

• BraTS 2021

• Challenge in MICCAI
• Evaluate state-of-the-art methods for the tumor segmentation in mpMRI scans

• Sliced 10 timesteps(70-79) to solve 2D Brain Tumor Segmentation task
• Sliced data have 240×240 resolution with 4 modalities (t1, t1ce, t2, flair), depending on whether a contrast agent is administered or not
• Each pixels are labeled one of 4 classes (label 0: Background, label 1: NCR, label 2: Edematous, label 4: ET)
• Preprocessed with min-max normalization to range each pixel values 0-1
• No data augmentation

• TBD

Final_Poster_DaehyeonChoi.pdf