UNet++ Compression Techniques for Kidney and Cyst Segmentation in Autosomal Dominant Polycystic Kidney Disease
-
- KRISHNAN Chetana
- Departments of Biomedical Engineering, The University of Alabama at Birmingham
-
- SCHMIDT Emma
- Departments of Biomedical Engineering, The University of Alabama at Birmingham
-
- ONUOHA Ezinwanne
- Departments of Biomedical Engineering, The University of Alabama at Birmingham
-
- MRUG Michal
- Departments of Nephrology, The University of Alabama at Birmingham Department of Veterans Affairs Medical Center
-
- CARDENAS Carlos E.
- Departments of Biomedical Engineering, The University of Alabama at Birmingham Departments of Radiation Oncology, The University of Alabama at Birmingham
-
- KIM Harrison
- Departments of Biomedical Engineering, The University of Alabama at Birmingham Departments of Radiology, The University of Alabama at Birmingham
抄録
<p>This study aimed to investigate the effect of UNet++ compression with pruning and principal component analysis (PCA) for kidney and cyst image segmentation of Autosomal Dominant Polycystic Kidney Disease (ADPKD). We used 756 T2-weighted MRI images of kidneys with ADPKD among the Consortium for Radiologic Imaging Studies of Polycystic Kidney Disease (CRISP) cohort. The UNet++, UNet++ with prune compression (prUNet++), and UNet++ with PCA compression (PCAUNet++) were trained, validated, and tested with 604, 76, and 76 kidney images, respectively. The model performance was analyzed using the Dice similarity coefficient (DSC). The training time per epoch for UNet++ was 217±5s and 220±7s respectively for kidneys and cysts. This was used as a pre-trained model for PCAUNet++ and prUNet++ whose training times after compression for kidneys and cysts respectively were 90±3s, 109±6s and, 110±2s, 124±4s. (p<0.0001). The test dataset’s average DSC values for UNet++, prUNet++, and PCAUNet++ were 0.93±0.05, 0.93±0.08, and 0.93±0.07, respectively, for kidney segmentation, while those for cyst segmentation were 0.87±0.04, 0.84±0.08, and 0.84±0.09, respectively, without statistical difference over the three models suggesting the pending rejection of the null hypothesis (p≥0.05). We demonstrated that compression techniques can be employed in segmentation tasks without the need to train a model from scratch and use an already trained model with additional little training, avoiding computational complexity and large training time. PCA-based compression technique significantly decreased the post-training time of UNet++ compared to prUNet++ without a significant reduction in the accuracy of kidney and cyst segmentation for ADPKD.</p>
収録刊行物
-
- Advanced Biomedical Engineering
-
Advanced Biomedical Engineering 13 (0), 134-143, 2024
公益社団法人 日本生体医工学会
- Tweet
キーワード
詳細情報 詳細情報について
-
- CRID
- 1390862500628461824
-
- ISSN
- 21875219
-
- 本文言語コード
- en
-
- データソース種別
-
- JaLC
- Crossref
-
- 抄録ライセンスフラグ
- 使用不可