Algorithm and Hardware Co-optimization for Image Segmentation in Wearable Ultrasound Devices: Continuous Bladder Monitoring
From monitoring muscle during exercise and training, to assess cardiovascular diseases, to estimate bladder volume, continuous autonomous tissue-monitoring is essential. Recent development in wearable ultrasound patches provide the foundation of wearable ultrasound devices with on-device image proce...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2023
|
| Online Access: | https://hdl.handle.net/1721.1/151618 |
| _version_ | 1826198750320656384 |
|---|---|
| author | Song, Zhiye |
| author2 | Chandrakasan, Anantha P. |
| author_facet | Chandrakasan, Anantha P. Song, Zhiye |
| author_sort | Song, Zhiye |
| collection | MIT |
| description | From monitoring muscle during exercise and training, to assess cardiovascular diseases, to estimate bladder volume, continuous autonomous tissue-monitoring is essential. Recent development in wearable ultrasound patches provide the foundation of wearable ultrasound devices with on-device image processing. Collaborating with Massachusetts General Hospital, we established bladder volume monitoring as the example use case. Real-time bladder monitoring can facilitate the diagnosis of post-operative urinary retention, and reduce indwelling urinary catheter usage and the risk of catheter-associated urinary tract infection. Using machine learning and hardware co-design, this thesis developed and validated a low-compute memory-efficient deep learning model and an energy-efficient all-parameters-on-chip application-specific integrated circuit (ASIC) for accurate bladder region segmentation and urine volume calculation.
U-Net is the state-of-the artneural network (NN) for biomedical image segmentation [1]. We trained two binarized models with 4-bits and 6-bits skip connections. They achieved an accuracy within 3.8% and 2.6% of the floating-point U-Net without any floating-point operations, and reduced memory requirement 11.5× and 9.0×, respectively, to under 150 kB. This thesis also designed the first neural network accelerator targeting U-Net-like image segmentation. Using interleaving feature map representation, skip connection compression, and extensive design space exploration, the accelerator does not require external memory or any co-processor, and consumes only 14.4μJ per 128 × 128 image segmentaiton.
The lightweight bladder volume estimation algorithm together with the energy-efficient image segmentation ASIC can be integrated with existing ultrasound probes to reduce the burdens of nurses in hospital settings and improve outpatient care. Moreover, the quantization and compression techniques and the image segmentation accelerator can be applied to other clinical applications, such as monitoring fetal heart rate and neural therapy. This technology, together with advances in compact ultrasound patches, will enable real-time tissue monitoring on the edge, thereby not only maintaining health data privacy, but also improving both point-of-care and inpatient healthcare. |
| first_indexed | 2024-09-23T11:09:31Z |
| format | Thesis |
| id | mit-1721.1/151618 |
| institution | Massachusetts Institute of Technology |
| last_indexed | 2024-09-23T11:09:31Z |
| publishDate | 2023 |
| publisher | Massachusetts Institute of Technology |
| record_format | dspace |
| spelling | mit-1721.1/1516182023-08-01T03:57:08Z Algorithm and Hardware Co-optimization for Image Segmentation in Wearable Ultrasound Devices: Continuous Bladder Monitoring Song, Zhiye Chandrakasan, Anantha P. Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science From monitoring muscle during exercise and training, to assess cardiovascular diseases, to estimate bladder volume, continuous autonomous tissue-monitoring is essential. Recent development in wearable ultrasound patches provide the foundation of wearable ultrasound devices with on-device image processing. Collaborating with Massachusetts General Hospital, we established bladder volume monitoring as the example use case. Real-time bladder monitoring can facilitate the diagnosis of post-operative urinary retention, and reduce indwelling urinary catheter usage and the risk of catheter-associated urinary tract infection. Using machine learning and hardware co-design, this thesis developed and validated a low-compute memory-efficient deep learning model and an energy-efficient all-parameters-on-chip application-specific integrated circuit (ASIC) for accurate bladder region segmentation and urine volume calculation. U-Net is the state-of-the artneural network (NN) for biomedical image segmentation [1]. We trained two binarized models with 4-bits and 6-bits skip connections. They achieved an accuracy within 3.8% and 2.6% of the floating-point U-Net without any floating-point operations, and reduced memory requirement 11.5× and 9.0×, respectively, to under 150 kB. This thesis also designed the first neural network accelerator targeting U-Net-like image segmentation. Using interleaving feature map representation, skip connection compression, and extensive design space exploration, the accelerator does not require external memory or any co-processor, and consumes only 14.4μJ per 128 × 128 image segmentaiton. The lightweight bladder volume estimation algorithm together with the energy-efficient image segmentation ASIC can be integrated with existing ultrasound probes to reduce the burdens of nurses in hospital settings and improve outpatient care. Moreover, the quantization and compression techniques and the image segmentation accelerator can be applied to other clinical applications, such as monitoring fetal heart rate and neural therapy. This technology, together with advances in compact ultrasound patches, will enable real-time tissue monitoring on the edge, thereby not only maintaining health data privacy, but also improving both point-of-care and inpatient healthcare. S.M. 2023-07-31T19:53:11Z 2023-07-31T19:53:11Z 2023-06 2023-07-13T14:29:41.628Z Thesis https://hdl.handle.net/1721.1/151618 In Copyright - Educational Use Permitted Copyright retained by author(s) https://rightsstatements.org/page/InC-EDU/1.0/ application/pdf Massachusetts Institute of Technology |
| spellingShingle | Song, Zhiye Algorithm and Hardware Co-optimization for Image Segmentation in Wearable Ultrasound Devices: Continuous Bladder Monitoring |
| title | Algorithm and Hardware Co-optimization for Image Segmentation in Wearable Ultrasound Devices: Continuous Bladder Monitoring |
| title_full | Algorithm and Hardware Co-optimization for Image Segmentation in Wearable Ultrasound Devices: Continuous Bladder Monitoring |
| title_fullStr | Algorithm and Hardware Co-optimization for Image Segmentation in Wearable Ultrasound Devices: Continuous Bladder Monitoring |
| title_full_unstemmed | Algorithm and Hardware Co-optimization for Image Segmentation in Wearable Ultrasound Devices: Continuous Bladder Monitoring |
| title_short | Algorithm and Hardware Co-optimization for Image Segmentation in Wearable Ultrasound Devices: Continuous Bladder Monitoring |
| title_sort | algorithm and hardware co optimization for image segmentation in wearable ultrasound devices continuous bladder monitoring |
| url | https://hdl.handle.net/1721.1/151618 |
| work_keys_str_mv | AT songzhiye algorithmandhardwarecooptimizationforimagesegmentationinwearableultrasounddevicescontinuousbladdermonitoring |