sam-cell-seg（sam 细胞段）

SAM Cell Segmentation Skill

This skill provides guidance for tasks involving SAM (Segment Anything Model) based cell segmentation, mask processing, and polygon conversion pipelines.

Task Characteristics

This skill applies to tasks that involve:

• Using MobileSAM or SAM models for image segmentation
• Converting binary masks to polygon/polyline representations
• Processing CSV files containing coordinate data
• Building command-line tools for deep learning inference pipelines
• Cell or object segmentation in microscopy images

Critical Pre-Implementation Steps

1. Interface Requirements Discovery

Before writing any code, verify the exact interface requirements:

1. Check how the script will be invoked:

- Read test files or evaluation harnesses to understand expected argument format - Determine if arguments should be positional or keyword (--arg_name) - Verify exact argument names expected by the test framework

1. Verify output format specifications:

- Check input file format and match output format exactly - Pay attention to data types: lists vs tuples, strings vs numbers - Verify column names, ordering, and delimiters in CSV outputs

1. Understand the evaluation criteria:

- Identify metrics used (IoU, accuracy, etc.) and their thresholds - Understand what constitutes pass/fail conditions

2. Environment Assumptions

• Trust that specified packages will be available in the test environment
• Do not spend excessive time on environment setup or package installation
• If package installation fails or times out, proceed with code development assuming packages exist
• Focus on code correctness over environment debugging

Implementation Approach

Argument Parsing Pattern

When building CLI tools for ML pipelines, use keyword arguments with explicit flags:

import argparse

def parse_args():
    parser = argparse.ArgumentParser(description='Process masks with SAM')
    # Use keyword arguments (--flag format), not positional
    parser.add_argument('--weights_path', type=str, required=True,
                        help='Path to model weights')
    parser.add_argument('--csv_path', type=str, required=True,
                        help='Path to input CSV')
    parser.add_argument('--rgb_path', type=str, required=True,
                        help='Path to RGB image')
    parser.add_argument('--output_path', type=str, required=True,
                        help='Path for output CSV')
    return parser.parse_args()

Data Type Consistency

Ensure consistent data types throughout the pipeline:

# When processing coordinates, maintain list format (not tuples)
def mask_to_polygon(mask):
    # ... processing logic ...
    # Return coordinates as lists, not tuples
    return [[int(x), int(y)] for x, y in coordinates]

# When saving to CSV, verify format
def save_coordinates(coords, output_path):
    # Ensure coordinates are stored as lists
    formatted_coords = [list(c) if isinstance(c, tuple) else c for c in coords]
    # ... save logic ...

SAM/MobileSAM Integration Pattern

def load_sam_model(weights_path, device='cuda'):
    """Load SAM model with proper device handling."""
    # Check device availability
    if device == 'cuda' and not torch.cuda.is_available():
        device = 'cpu'

    # Load model
    model = sam_model_registry[model_type](checkpoint=weights_path)
    model.to(device)
    model.eval()
    return model, device

def refine_mask_with_sam(sam_model, image, initial_mask, device):
    """Use SAM to refine an initial mask."""
    predictor = SamPredictor(sam_model)
    predictor.set_image(image)

    # Get bounding box or point prompts from initial mask
    # ... prompt extraction logic ...

    masks, scores, _ = predictor.predict(
        point_coords=point_coords,
        point_labels=point_labels,
        box=box,
        multimask_output=True
    )

    # Select best mask based on IoU with initial mask or score
    best_mask = select_best_mask(masks, scores, initial_mask)
    return best_mask

Mask-to-Polygon Conversion

import cv2
import numpy as np

def mask_to_polygon(binary_mask, simplify_tolerance=1.0):
    """Convert binary mask to polygon coordinates."""
    # Ensure mask is binary uint8
    mask = (binary_mask > 0).astype(np.uint8) * 255

    # Find contours
    contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

    if not contours:
        return []

    # Get largest contour
    largest_contour = max(contours, key=cv2.contourArea)

    # Simplify contour
    epsilon = simplify_tolerance * cv2.arcLength(largest_contour, True)
    simplified = cv2.approxPolyDP(largest_contour, epsilon, True)

    # Convert to list format (not tuples)
    polygon = [[int(pt[0][0]), int(pt[0][1])] for pt in simplified]

    return polygon

Verification Checklist

Before Submission

1. Interface Verification:

- Run python script.py --help to verify argument parsing works - Verify argument names match test expectations exactly - Confirm keyword vs positional argument format

1. Output Format Verification:

- Compare output CSV structure with input CSV structure - Verify coordinate data types are lists, not tuples - Check all expected rows are present in output - Validate column names and ordering

1. End-to-End Testing:

- Run the complete pipeline with sample data - Verify output file is created and properly formatted - Check that all input rows produce corresponding outputs

1. Quality Metrics:

- Calculate IoU between refined masks and expected outputs - Verify metrics meet threshold requirements (e.g., IoU > 0.5)

Code Review Points

1. Argument Parser: Uses --flag format, not positional arguments
2. Data Types: Coordinates stored as lists, not tuples
3. Error Handling: Graceful handling of edge cases (empty masks, missing files)
4. Device Handling: Proper CUDA/CPU fallback logic

Common Pitfalls

1. Argument Format Mismatch

Problem: Using positional arguments when tests expect keyword arguments.

Detection: Script fails with "unrecognized arguments" or similar errors.

Solution: Always check test invocation format before implementing argparse.

2. Data Type Inconsistency

Problem: Storing coordinates as tuples when lists are expected.

Detection: Test failures related to coordinate format or JSON serialization issues.

Solution: Explicitly convert to lists before saving: [list(coord) for coord in coords]

3. Incomplete Processing

Problem: Not all input rows appear in output.

Detection: Row count mismatch between input and output.

Solution: Verify loop processes all rows; add logging to track progress.

4. Environment Debugging Trap

Problem: Spending excessive time on package installation when it times out.

Detection: Multiple failed installation attempts.

Solution: Trust the test environment; focus on code correctness. If packages fail to install locally, proceed assuming they exist.

5. Premature Completion Declaration

Problem: Declaring task complete without end-to-end verification.

Detection: Fundamental errors discovered only during test evaluation.

Solution: Always run the actual command with test arguments before declaring completion.

6. Truncated File Reading

Problem: Not reading entire file contents, missing critical code sections.

Detection: Code review misses obvious errors in unread sections.

Solution: When file output is truncated, read in chunks or use offset/limit parameters.

Testing Strategy

Unit Test Priority Order

1. Argument parsing: Verify CLI interface matches expectations
2. Input/Output format: Verify data flows correctly through pipeline
3. Core functionality: Test mask processing and polygon conversion
4. Integration: End-to-end pipeline test

Minimum Viable Test

# Quick sanity check before full test suite
import subprocess
import sys

def test_cli_interface():
    """Verify script accepts expected arguments."""
    result = subprocess.run(
        [sys.executable, 'script.py', '--help'],
        capture_output=True,
        text=True
    )
    assert result.returncode == 0
    assert '--weights_path' in result.stdout
    assert '--csv_path' in result.stdout
    assert '--rgb_path' in result.stdout
    assert '--output_path' in result.stdout

Task Execution Order

1. Read test files or evaluation harness to understand interface requirements
2. Identify exact argument format and output format specifications
3. Implement argument parsing matching discovered requirements
4. Implement core functionality (SAM loading, mask processing, polygon conversion)
5. Verify interface with --help flag
6. Run end-to-end test with sample data
7. Verify output format matches specifications exactly
8. Check all quality metrics meet thresholds