{ "id": 672944, "title": "51st place solution", "url": "/competitions/physionet-ecg-image-digitization/discussion/672944", "author_name": "COLA KE LIU", "post_date": "2026-02-11T11:15:08.544000", "votes": 1, "comment_count": 0, "views": 0, "content": "

3-Stage Pipeline with Robust Signal Extraction

\n

First of all, thanks to the organizers for hosting this rerun of the ECG digitization challenge. Here is an overview of our team's approach.

\n

Overview

\n

Our solution breaks the problem down into three distinct sub-tasks:

\n
    \n
  1. Stage 0: ROI Extraction & Orientation: Locating the ECG chart and correcting perspective.
    2. \n
  2. Stage 1: Rectification: Unwarping the image to correct non-linear paper deformations.
    3. \n
  3. Stage 2: Segmentation: Predicting the probability map of the signal trace.
    4. \n
\n

Key Improvement: The critical difference in our final submission lies in our Signal Extraction (Post-processing) strategy (fill_pro.py), where we replaced simple peak detection with Gaussian smoothing and advanced interpolation to handle broken signal lines and noise.

\n
\n

1. Preprocessing & Source Classification

\n

We noticed that the test set contains images from various sources (scanners, photos, different resolutions). Applying the same preprocessing to all images hurts performance.

\n\n

2. The Pipeline Architecture

\n

Stage 0: Global Alignment (Homography)

\n\n

Stage 1: Fine-Grained Rectification

\n\n

Stage 2: 1D Signal Segmentation

\n\n
\n

3. The \"Secret Sauce\": Robust Signal Extraction (fill_pro)

\n

In the baseline approach, extracting the time series from the Stage 2 heatmap was done via a simple argmax. However, this fails when:

\n
    \n
  1. The ink is faded (missing signal).
    2. \n
  2. There is noise (salt-and-pepper artifacts).
    3. \n
  3. The line is jagged.
    4. \n
\n

We implemented a robust extraction algorithm (fill_pro.py) that significantly improved our SNR score.

\n

A. Gaussian Probability Smoothing

\n

Before extracting the coordinates, we apply a Gaussian Blur to the probability map output by the model.

\n\n

B. Handling Signal Gaps (Interpolation vs. Zero-Filling)

\n

The baseline approach filled missing data (where model confidence < threshold) with the zero_mv (isoelectric line). This causes massive penalties in the SNR metric because a missing peak is treated as a flat line (high error).

\n

We replaced this with Linear/Spline Interpolation.

\n\n

C. Einthoven's Law Correction

\n

We apply a physics-based constraint for the limb leads. According to Einthoven's Law, Lead II should equal Lead I + Lead III. We calculate the error between the measured leads and distribute this error to correct the signals.

\n

This post-processing step ensures the extracted signals are electrically consistent, acting as a powerful regularizer against noise in any single lead.

\n
\n

4. Summary of Parameters

\n

For the final submission, we used the following configuration in fill_pro:

\n", "messages": [ { "id": 3404855, "postDate": "2026-02-11T11:15:08.543Z", "content": "

3-Stage Pipeline with Robust Signal Extraction

\n

First of all, thanks to the organizers for hosting this rerun of the ECG digitization challenge. Here is an overview of our team's approach.

\n

Overview

\n

Our solution breaks the problem down into three distinct sub-tasks:

\n
    \n
  1. Stage 0: ROI Extraction & Orientation: Locating the ECG chart and correcting perspective.
    2. \n
  2. Stage 1: Rectification: Unwarping the image to correct non-linear paper deformations.
    3. \n
  3. Stage 2: Segmentation: Predicting the probability map of the signal trace.
    4. \n
\n

Key Improvement: The critical difference in our final submission lies in our Signal Extraction (Post-processing) strategy (fill_pro.py), where we replaced simple peak detection with Gaussian smoothing and advanced interpolation to handle broken signal lines and noise.

\n
\n

1. Preprocessing & Source Classification

\n

We noticed that the test set contains images from various sources (scanners, photos, different resolutions). Applying the same preprocessing to all images hurts performance.

\n\n

2. The Pipeline Architecture

\n

Stage 0: Global Alignment (Homography)

\n\n

Stage 1: Fine-Grained Rectification

\n\n

Stage 2: 1D Signal Segmentation

\n\n
\n

3. The \"Secret Sauce\": Robust Signal Extraction (fill_pro)

\n

In the baseline approach, extracting the time series from the Stage 2 heatmap was done via a simple argmax. However, this fails when:

\n
    \n
  1. The ink is faded (missing signal).
    2. \n
  2. There is noise (salt-and-pepper artifacts).
    3. \n
  3. The line is jagged.
    4. \n
\n

We implemented a robust extraction algorithm (fill_pro.py) that significantly improved our SNR score.

\n

A. Gaussian Probability Smoothing

\n

Before extracting the coordinates, we apply a Gaussian Blur to the probability map output by the model.

\n\n

B. Handling Signal Gaps (Interpolation vs. Zero-Filling)

\n

The baseline approach filled missing data (where model confidence < threshold) with the zero_mv (isoelectric line). This causes massive penalties in the SNR metric because a missing peak is treated as a flat line (high error).

\n

We replaced this with Linear/Spline Interpolation.

\n\n

C. Einthoven's Law Correction

\n

We apply a physics-based constraint for the limb leads. According to Einthoven's Law, Lead II should equal Lead I + Lead III. We calculate the error between the measured leads and distribute this error to correct the signals.

\n

This post-processing step ensures the extracted signals are electrically consistent, acting as a powerful regularizer against noise in any single lead.

\n
\n

4. Summary of Parameters

\n

For the final submission, we used the following configuration in fill_pro:

\n", "rawMarkdown": "# 3-Stage Pipeline with Robust Signal Extraction\n\nFirst of all, thanks to the organizers for hosting this rerun of the ECG digitization challenge. Here is an overview of our team's approach.\n\n## Overview\nOur solution breaks the problem down into three distinct sub-tasks:\n1. **Stage 0: ROI Extraction & Orientation:** Locating the ECG chart and correcting perspective.\n2. **Stage 1: Rectification:** Unwarping the image to correct non-linear paper deformations.\n3. **Stage 2: Segmentation:** Predicting the probability map of the signal trace.\n\n**Key Improvement:** The critical difference in our final submission lies in our **Signal Extraction (Post-processing)** strategy (`fill_pro.py`), where we replaced simple peak detection with Gaussian smoothing and advanced interpolation to handle broken signal lines and noise.\n\n---\n\n## 1. Preprocessing & Source Classification\nWe noticed that the test set contains images from various sources (scanners, photos, different resolutions). Applying the same preprocessing to all images hurts performance.\n\n* **Source Classifier:** We trained an `EfficientNet-B2` to classify the image \"source\" (12 classes).\n* **Adaptive Preprocessing:** Based on the predicted source ID, we apply specific OpenCV operations:\n * **CLAHE:** For low-contrast images.\n * **Gray-world White Balance:** For photos with color casts.\n * **Bilateral/Median Filtering:** For noisy scans.\n * **Background Removal:** Using morphological opening (Lab color space) to remove uneven lighting.\n\n## 2. The Pipeline Architecture\n\n### Stage 0: Global Alignment (Homography)\n* **Model:** ResNet34 Encoder.\n* **Task:** Predicts keypoints (corners/anchors) and orientation.\n* **Action:** We calculate a Homography matrix to warp the input image into a standard coordinate system. This handles rotation and perspective distortion.\n\n### Stage 1: Fine-Grained Rectification\n* **Model:** U-Net style architecture.\n* **Task:** Detects the grid lines (horizontal and vertical) and grid intersections.\n* **Action:** We use the predicted grid points to create a dense flow map. Using `torch.nn.functional.grid_sample`, we \"unwarp\" the image. This is crucial for photos of paper that is curved or folded, ensuring the time-axis is linear.\n\n### Stage 2: 1D Signal Segmentation\n* **Model:** ResNet34 Encoder + `CoordUnetDecoder`.\n* **Feature:** We inject Coordinate Convolutions (`CoordConv`) in the decoder to help the model understand the spatial position of the signals (since ECG leads have fixed positions).\n* **Output:** A probability map (Heatmap) of the signal trace.\n\n---\n\n## 3. The \"Secret Sauce\": Robust Signal Extraction (`fill_pro`)\n\nIn the baseline approach, extracting the time series from the Stage 2 heatmap was done via a simple `argmax`. However, this fails when:\n1. The ink is faded (missing signal).\n2. There is noise (salt-and-pepper artifacts).\n3. The line is jagged.\n\nWe implemented a robust extraction algorithm (`fill_pro.py`) that significantly improved our SNR score.\n\n### A. Gaussian Probability Smoothing\nBefore extracting the coordinates, we apply a Gaussian Blur to the probability map output by the model.\n\n* **Why:** The raw pixel predictions can be noisy. Blurring aggregates local evidence, making the `argmax` (peak detection) much more stable and continuous.\n* **Parameters:** We tuned `sigma=0.29` and `threshold=0.11` for the best leaderboard performance.\n\n### B. Handling Signal Gaps (Interpolation vs. Zero-Filling)\nThe baseline approach filled missing data (where model confidence < threshold) with the `zero_mv` (isoelectric line). This causes massive penalties in the SNR metric because a missing peak is treated as a flat line (high error).\n\nWe replaced this with **Linear/Spline Interpolation**.\n\n* **Impact:** If a QRS complex has a gap in the middle due to poor scanning, interpolation connects the dots, preserving the shape much better than dropping to zero.\n\n### C. Einthoven's Law Correction\nWe apply a physics-based constraint for the limb leads. According to Einthoven's Law, Lead II should equal Lead I + Lead III. We calculate the error between the measured leads and distribute this error to correct the signals.\n\nThis post-processing step ensures the extracted signals are electrically consistent, acting as a powerful regularizer against noise in any single lead.\n\n---\n\n## 4. Summary of Parameters\nFor the final submission, we used the following configuration in `fill_pro`:\n* **Smoothing:** Gaussian (`sigma=0.29`).\n* **Interpolation:** Linear (`scipy.interpolate.interp1d`).\n* **Threshold:** `0.11` (Pixels with probability below this are treated as missing and interpolated).\n* **Post-Smoothing:** Savitzky-Golay filter (`window=7`, `poly=2`) applied to the final 1D series to remove high-frequency digitization noise.", "votes": 1 } ], "comments": [], "raw_markdown_by_id": { "3404855": "# 3-Stage Pipeline with Robust Signal Extraction\n\nFirst of all, thanks to the organizers for hosting this rerun of the ECG digitization challenge. Here is an overview of our team's approach.\n\n## Overview\nOur solution breaks the problem down into three distinct sub-tasks:\n1. **Stage 0: ROI Extraction & Orientation:** Locating the ECG chart and correcting perspective.\n2. **Stage 1: Rectification:** Unwarping the image to correct non-linear paper deformations.\n3. **Stage 2: Segmentation:** Predicting the probability map of the signal trace.\n\n**Key Improvement:** The critical difference in our final submission lies in our **Signal Extraction (Post-processing)** strategy (`fill_pro.py`), where we replaced simple peak detection with Gaussian smoothing and advanced interpolation to handle broken signal lines and noise.\n\n---\n\n## 1. Preprocessing & Source Classification\nWe noticed that the test set contains images from various sources (scanners, photos, different resolutions). Applying the same preprocessing to all images hurts performance.\n\n* **Source Classifier:** We trained an `EfficientNet-B2` to classify the image \"source\" (12 classes).\n* **Adaptive Preprocessing:** Based on the predicted source ID, we apply specific OpenCV operations:\n * **CLAHE:** For low-contrast images.\n * **Gray-world White Balance:** For photos with color casts.\n * **Bilateral/Median Filtering:** For noisy scans.\n * **Background Removal:** Using morphological opening (Lab color space) to remove uneven lighting.\n\n## 2. The Pipeline Architecture\n\n### Stage 0: Global Alignment (Homography)\n* **Model:** ResNet34 Encoder.\n* **Task:** Predicts keypoints (corners/anchors) and orientation.\n* **Action:** We calculate a Homography matrix to warp the input image into a standard coordinate system. This handles rotation and perspective distortion.\n\n### Stage 1: Fine-Grained Rectification\n* **Model:** U-Net style architecture.\n* **Task:** Detects the grid lines (horizontal and vertical) and grid intersections.\n* **Action:** We use the predicted grid points to create a dense flow map. Using `torch.nn.functional.grid_sample`, we \"unwarp\" the image. This is crucial for photos of paper that is curved or folded, ensuring the time-axis is linear.\n\n### Stage 2: 1D Signal Segmentation\n* **Model:** ResNet34 Encoder + `CoordUnetDecoder`.\n* **Feature:** We inject Coordinate Convolutions (`CoordConv`) in the decoder to help the model understand the spatial position of the signals (since ECG leads have fixed positions).\n* **Output:** A probability map (Heatmap) of the signal trace.\n\n---\n\n## 3. The \"Secret Sauce\": Robust Signal Extraction (`fill_pro`)\n\nIn the baseline approach, extracting the time series from the Stage 2 heatmap was done via a simple `argmax`. However, this fails when:\n1. The ink is faded (missing signal).\n2. There is noise (salt-and-pepper artifacts).\n3. The line is jagged.\n\nWe implemented a robust extraction algorithm (`fill_pro.py`) that significantly improved our SNR score.\n\n### A. Gaussian Probability Smoothing\nBefore extracting the coordinates, we apply a Gaussian Blur to the probability map output by the model.\n\n* **Why:** The raw pixel predictions can be noisy. Blurring aggregates local evidence, making the `argmax` (peak detection) much more stable and continuous.\n* **Parameters:** We tuned `sigma=0.29` and `threshold=0.11` for the best leaderboard performance.\n\n### B. Handling Signal Gaps (Interpolation vs. Zero-Filling)\nThe baseline approach filled missing data (where model confidence < threshold) with the `zero_mv` (isoelectric line). This causes massive penalties in the SNR metric because a missing peak is treated as a flat line (high error).\n\nWe replaced this with **Linear/Spline Interpolation**.\n\n* **Impact:** If a QRS complex has a gap in the middle due to poor scanning, interpolation connects the dots, preserving the shape much better than dropping to zero.\n\n### C. Einthoven's Law Correction\nWe apply a physics-based constraint for the limb leads. According to Einthoven's Law, Lead II should equal Lead I + Lead III. We calculate the error between the measured leads and distribute this error to correct the signals.\n\nThis post-processing step ensures the extracted signals are electrically consistent, acting as a powerful regularizer against noise in any single lead.\n\n---\n\n## 4. Summary of Parameters\nFor the final submission, we used the following configuration in `fill_pro`:\n* **Smoothing:** Gaussian (`sigma=0.29`).\n* **Interpolation:** Linear (`scipy.interpolate.interp1d`).\n* **Threshold:** `0.11` (Pixels with probability below this are treated as missing and interpolated).\n* **Post-Smoothing:** Savitzky-Golay filter (`window=7`, `poly=2`) applied to the final 1D series to remove high-frequency digitization noise." } }