Machine Intelligence for Accurate X-ray Screening and Read-out Prioritization: PICC Line Detection Study

Transcription

1 Machine Intelligence for Accurate X-ray Screening and Read-out Prioritization: PICC Line Detection Study Laboratory of Medical Imaging and Computation Massachusetts General Hospital Hyunkwang Lee, Jordan Rogers Junghwan Cho PhD, Dania Daye MD, Vishala Mishra MD, Garry Choy MD, Shahein Tajmir MD, Michael Lev MD, Synho Do PhD

2 Disclosure Thank you for SIIM 2017 Annual Meeting New Investigator Travel Award

3 Introduction Peripherally inserted central catheter (PICC) for intravenous access Malpositioned PICCs serious complications Final PICC location confirmed using a chest radiograph High accuracy by humans, but delays in interpretation PICC

4 PICC Detection System Test PICC Detection AI Train Input data Predict Postprocessing Preprocessing #SIIM17 Test image Train images Final output image

5 PICC Detection System Test PICC Detection AI Train Input data Predict Postprocessing Preprocessing #SIIM17 Test image Train images Final output image

6 PICC Detection System Test PICC Detection AI Train Input data Predict Postprocessing Preprocessing #SIIM17 Test image Train images Final output image

7 PICC Detection System Test PICC Detection AI Train Input data Predict Postprocessing Preprocessing #SIIM17 Test image Train images Final output image

8 Dataset (Collection) 800 DICOM images 600 for train and validation, 200 for test

9 Dataset (Characteristics) Low contrast High contrast Artifacts Patient rotation Low image quality High variance in pixel contrast, artifacts, patient rotation High resolution, various size (2801x3195 pixels on average)

10 PICC Detection System Test PICC Detection AI Train Input data Predict Postprocessing Preprocessing #SIIM17 Test image Train images Final output image

11 Preprocessing Preprocessing module Original Preprocessed Bilateral filter for denoising and edge enhancement Adaptive Histogram Equalization for image contrast normalization

12 PICC Detection System Test PICC Detection AI (Patch-based Approach) Train Input data Predict Postprocessing Preprocessing #SIIM17 Test image Train images Final output image

13 PICC Detection AI (Patch-based Approach) 1) Input data Manual annotations for 10 classes PICC, ECG Tissue, Rib, Vertebral body, Shoulder, Lung Other lines, Other objects, Background

14 PICC Detection AI (Patch-based Approach) 2) Train Mirroring Rotation (-90 to 90 ) Sampling patches Data Augmentation 10 classes 96 pixels square Balanced dataset (70K images / class) Fine-tuned, ImageNet pretrained AlexNet Model selection on validation results

15 PICC Detection AI (Patch-based Approach) 3) Test Execute CNN Per-patch classification results Preprocessed test image Sampled Patches Predicted PICC mask 96 pixels square, stride 12 Avg. 58,050 patches/image

16 PICC Detection System Test PICC Detection AI Train Input data Predict Postprocessing Preprocessing #SIIM17 Test image Train images Final output image

17 Postprocessing Postprocessing module PICC Tip Predicted PICC mask Used a Hough line transform algorithm To prune false positives To merge the significant nearby contours Postprocessed PICC mask Final output

18 Result

19 Result Absolute Distance Mean ± Std RMSE 4.66 ± 2.8 mm 5.44 mm

20 Result Absolute Distance Mean ± Std RMSE 4.66 ± 2.8 mm 5.44 mm Execution Time sec/image on a GPU Redundant pixel-wise computations #SIIM17 Avg. 58,050 patches/image

21 PICC Detection AI (Fully Convolutional Network) Test PICC Detection AI (Fully Convolutional Network) Train Input data Predict Postprocessing Preprocessing #SIIM17 Test image Train image Final output

22 PICC Detection AI (Fully Convolutional Network) Test Fully Convolutional Network Predict Postprocessing Test image Train Predicted PICC mask PICC Tip Images Train dataset Labels Final output

23 Final Results Absolute distance Mean ± Std RMSE Patch-based Approach 4.66 ± 2.8 mm 5.44 mm Execution time Mean sec/image on a GPU

24 Final Results Patch-based Approach Fully convolutional Network Absolute distance Mean ± Std 4.66 ± 2.8 mm 3.10 ± 2.03 mm RMSE 5.44 mm 3.71 mm Deployment time Mean sec/image on a GPU 1.32 sec/image on a GPU 1.5x better performance 66.4x faster execution time

25 Other Functions Non-existent PICC Partially visualized PICC Normal PICC To identify non-existent or partially visualized PICC cases

26 Other Functions PICC ECG Objects Threads Automatic highlights on medical devices. Generalized to other types of vascular access and therapeutic support devices

27 Conclusion Results Performance : 3.10 ± 2.03 mm Execution Time : 1.32 sec/image on a single GPU Clinical application Implemented on the imaging device or as part of PACS for triage correct vs incorrect Future work Extension to detect other types of vascular access and therapeutic support devices

28 Thank you for attention Hyunkwang Lee hyunkwanglee@seas.harvard.edu Homepage : scholar.harvard.edu/hklee

29