Breast Cancer Ultrasound Segmentation
9-architecture segmentation benchmark on 780 BUSI images. DeepLabV3+ tops with Dice 0.7863, IoU 0.6483. FCN → SimpleUNet → SegNet → Attention-UNet → TransUNet → ResNet34-UNet → EfficientNet-UNet → DeepLabV3+ → Swin-UNet.
780 breast ultrasound images (437 benign, 210 malignant, 133 normal)
9-architecture benchmark: classical encoders → transformer (TransUNet) → ASPP (DeepLabV3+)
Comprehensive benchmark of 9 segmentation architectures on the Breast Ultrasound Images (BUSI) dataset.
Dataset
- ▸780 images: 437 benign / 210 malignant / 133 normal (class-stratified)
- ▸Split: 585 train / 117 val / 78 test
- ▸Preprocessing: 256×256 resize, Albumentations augmentation (elastic, brightness/contrast, horizontal flip)
- ▸Loss: 50/50 BCE + Dice combined
Full Architecture Benchmark
| Model | Params | Dice | IoU | Type |
|---|---|---|---|---|
| FCN | — | 0.4822 | — | Semantic baseline |
| SimpleUNet | 31M | 0.7030 | 0.5424 | Standard encoder-decoder |
| SegNet | 7M | 0.6725 | — | Pooling-index decoder |
| Attention UNet | 31M | 0.6964 | — | Attention gates on skip |
| TransUNet | 124M | 0.6663 | — | ViT encoder + CNN decoder |
| EfficientNet-UNet | 20M | 0.7241 | 0.5683 | EfficientNet-B4 encoder |
| ResNet34-UNet | 24M | 0.7837 | 0.6440 | Pretrained ResNet encoder |
| DeepLabV3+ | 27M | 0.7863 | 0.6483 | ASPP multi-scale |
| Swin-UNet | High | — | — | Transformer-based |
Training Setup
- ▸30 epochs, AdamW + cosine annealing LR
- ▸Mixed precision (AMP), early stopping
- ▸Metrics: Dice, IoU, F1, Precision, Recall (TorchMetrics)
Why DeepLabV3+ Wins Atrous Spatial Pyramid Pooling (ASPP) with output_stride=16 extracts multi-scale context without spatial collapse — critical for breast tumors that range from 2mm to 30mm. This prevents the spatial resolution loss that hurts standard encoders on small lesions.