Basars Model

Params

Pre-trained encoder: R50+ViT-B/16
Input shape: (224, 224, 3)
Encoder trainable: False
Batch size: 64
Epochs: 2,190

Model

Input Image (224, 224, 3) → Vision Transformer + CUP (224, 224, 16) → Conv1x1 (224, 224, 5)

Total params: 100,861,024
Trainable params: 7,409,632
Non-trainable params: 93,451,392

Dataset

• Medial Images for Gastric Cancer Diagnosis (Korean)

Preprocessing

• Preprocessors

1. Remove malformed images from entire dataset.
2. Fill missing polygons from label data using CVAT. (Thanks to @kdh93)
3. Exports its ROI, binary masks and transformed original images.

Augmentation

Used tf.data.Dataset to boost training performance

• Scaling (1.0 / 255)
• Random Flip Left-Right
• Random Flip Up-Down
• Random Crop
• Random Brightness (-0.2 ~ +0.2)
• Random Rotation (90°, 180°, 270°)
• Gaussian Noise (mean = 0, stddev = 0.05)

Compile options

⚠️ The Reasons of Bad Results

The results are hilariously bad, because we didn't know that we must use Focal Loss or Weighted CCE for imbalanced dataset. Although we no longer have access to the GPU server, now we know that we could have better results.

Loss

• CCE * 0.5 + Dice * 0.5

Metrics

• Binary IoU
  • Number of classes = 5 (default)
  • Threshold = 0.5
  • Masked with [0, 1, 1, 1, 1] to ignore non-cancer area

Optimizer

• SGD
  • Momentum = 0.9
  • Learning Rate Scheduler
    • Cosine Annealing Warmup Restarts
      • First cycle steps = 100
      • Initial learning rate = 1e-3
      • First decay steps = 300
      • t_mul = 1.0
      • m_mul = 1.0 (default)
      • alpha = 0.0 (default)

Cosine Annealing Warmup Restarts:

Results

Evaluation

Segmentation Performance

Classification Performance