Matryoshka Representation Learning for Medical Image Segmentation
MD. Abir Hasan Piash
Undergraduate Thesis · Dept. of CSE, BRAC University
Abstract
MatSeg-Conv is an encoder-decoder network that emits coarse, medium and fine segmentation masks from a single forward pass through nested 16/64/256-channel capacity bands, so one trained model serves every hardware budget instead of retraining a family of them.
Keywords - PyTorch · ConvNeXt · ASPP · Deep Supervision · Medical Imaging
- 1.Combined a ConvNeXt-Tiny encoder, a nested feature interface at every encoder stage, Atrous Spatial Pyramid Pooling and three capacity-specific decoder heads trained under deep supervision (weighted BCE + Dice) in roughly 29M parameters.
- 2.Beat U-Net, Attention U-Net, ResUNet and DeepLabV3+ (31M–39M parameters) on cross-dataset mean Dice while remaining the smallest model benchmarked, across brain MRI, colonoscopy, electron microscopy and dermoscopy.
- 3.Showed strong low-data robustness: up to +0.16 Dice over DeepLabV3+ on 50-image training subsets, improving monotonically across every reduced-data setting and tumour category.
- 4.Ran a controlled single-variable ablation across ten architectural configurations, retaining only the modifications that improved validation performance.
| Measure | Value |
|---|---|
| Mean Dice | 0.8326 |
| Mean IoU | 0.7488 |
| Parameters | 29M |
| Imaging modalities | 4 |
