Embedded Files
Figure 1: Enhanced performance of vertebral body detection after applying the proposed novel windowing algorithm
DICOM Windowing Algorithm
DICOM Windowing Algorithm
Enhanced the current windowing algorithm to handle DICOM images in a better way. The current algorithm was limited due to (1) MONOCHROME issues, as well as (2) incorrectly inputted DICOM tag values.
MONOCHROME Issues
Simple conversion between MONOCHROME1 --> MONOCHROME2 wouldn't work.
Python SITK package's built-in MONOCHROME1 --> MONOCHROME2 automatic conversion wouldn't work for all cases. It converts the image itself, but not the 4 core DICOM tag values along with it.
Incorrectly inputted DICOM tag values
There are meaningless values in slope and intercept parameters in DICOM files which makes the automatic windowing to be failed.
DICOM tags are depending on how the institutions set up their machines to capture medical image, and they inputted those values, which is out of researcher's control.
Technical limitations: Python package SITK results different DICOM tag values from Microdicom GUI SW, since some DICOM files have values under same tag, it perhaps confused the SITK package.
The proposed novel windowing algorithm handles them by calculating the window center and width values with the bits stored value. Transformation in pixel distribution also helps when it gets fed into the CNN-based ML models. Yet, it is limited to outliers especially when the DICOM image itself has an inputted text to mark "L" (left of body) and etc.
Spine CNN Classification Models
Spine CNN Classification Models
Multiple CNN-based machine learning models have developed as well, to classify whether the input DICOM spinal medical image should be analyzed for compression fracture or scoliosis calculations. The following classification models have been built:
View Model: Anteroposterior (AP), Lateral (LAT), Oblique (OBL), Cone View
Posture Model: Plain, Bending
Exception Model: Assist, Gas, Gas Only, Spine
Accuracy Overview:
View Model: 0.9914
Posture Model: 0.9692
Exception Model: 0.8563
News
News
[ Nov 2021 ] The proposed classification model is in an installation process to the multiple affiliated hospitals in nationwide (South Korea), including Seoul National University Hospital (SNUH).
[ Dec 2022 ] The proposed DICOM file windowing algorithm and the classification model are adapted to the deep-learning-based algorithm that measures the compression ratio (CR) for compression fracture (CF) diagnosis in lateral X-rays. This work was presented at Radiological Society of North America (RSNA 2022).
Radiological Society of North America (RSNA 2022) Abstract
Radiological Society of North America (RSNA 2022) Abstract
Purpose: To develop a deep-learning-based algorithm that conducts reliable measuring of compression ratio (CR) for compression fracture (CF) diagnosis in lateral X-rays.
Methods and Materials: We, retrospectively, collected 7,650 spine X-ray images which has been sent to the government agency under ministry of health and welfare responsible for claims review and quality assessment of the Health Insurance. We used four classification models, view, body part, posture, and assistant device, to classify whether the input radiography is a subject for CF analysis. With U-Net model, trained with 1,040 lateral TL-spine X-rays and annotations, we segment vertebra body (VB) to calculate the compression ratio by each subject's front height over average of the two adjacent ones. We calculated CR by the subject VB’s front height over average of the two adjacent VB’s front heights.
Results: For evaluating performance of our algorithm, we used 250 spine X-ray images, included 145 TL-Spine X-ray images suitable for CR measurment (44.18%±13.04%, min: 20%, max: 74.72%). The sequential CF subject classification model show an accuracy of 94.8% (95% CI: 0.912≤p≤0.972). Mean Absolute Error (MAE) between ground truth and the predicted CR was 6.04% with standard deviation of 8.86%. By evaluating the above performance by ground truth’s severity, the mild, moderate, and, severe fracture performances were 6.54%±7.64%, 4%±3.78% and 7.2%±10.74% respectively.
Conclusions: In practice, it is essential to check whether the input image is suitable for specific analysis algorithms. Unsuitable input could give false result, potentially misguide medical practitioner, and even slow down the process, which is the opposite of the purpose of automated system. Result of this study shows that our automatic algorithm could not only provide an efficient change in process of measuring CR with high performance, but also be an objective solution by selecting subject analysis images.
Clinical Relevance/Application: To apply deep learning based automated analysis algorithm into actual practice, we propose a sequential classification model to select the subject image for analysis of CF, which is required to be plain lateral TL-spine x-ray image.
Page updated
Google Sites
Report abuse