Efficient knowledge distillation for liver CT segmentation using growing assistant network

Knowledge distillation on individual vertebrae segmentation exploiting 3D U-Net

Recent advances in medical imaging have highlighted the critical development of algorithms for individual vertebral segmentation on computed tomography (CT) scans. Essential for diagnostic accuracy and treatment planning in orthopaedics, neurosurgery and oncology, these algorithms face challenges in clinical implementation, including integration into healthcare systems. Consequently, our focus lies in exploring the application of knowledge distillation (KD) methods to train shallower networks capable of efficiently segmenting vertebrae in CT scans. This approach aims to reduce segmentation time, enhance suitability for emergency cases, and optimize computational and memory resource efficiency. Building upon prior research in the field, a two-step segmentation approach was employed. Firstly, the spine's location was determined by predicting a heatmap, indicating the probability of each voxel belonging to the spine. Subsequently, an iterative segmentation of vertebrae was performed from the top to the bottom of the CT volume over the located spine, using a memory instance to record the already segmented vertebrae. KD methods were implemented by training a teacher network with performance similar to that found in the literature, and this knowledge was distilled to a shallower network (student). Two KD methods were applied: (1) using the soft outputs of both networks and (2) matching logits. Two publicly available datasets, comprising 319 CT scans from 300 patients and a total of 611 cervical, 2387 thoracic, and 1507 lumbar vertebrae, were used. To ensure dataset balance and robustness, effective data augmentation methods were applied, including cleaning the memory instance to replicate the first vertebra segmentation. The teacher network achieved an average Dice similarity coefficient (DSC) of 88.22% and a Hausdorff distance (HD) of 7.71 mm, showcasing performance similar to other approaches in the literature. Through knowledge distillation from the teacher network, the student network's performance improved, with an average DSC increasing from 75.78% to 84.70% and an HD decreasing from 15.17 mm to 8.08 mm. Compared to other methods, our teacher network exhibited up to 99.09% fewer parameters, 90.02% faster inference time, 88.46% shorter total segmentation time, and 89.36% less associated carbon (CO2) emission rate. Regarding our student network, it featured 75.00% fewer parameters than our teacher, resulting in a 36.15% reduction in inference time, a 33.33% decrease in total segmentation time, and a 42.96% reduction in CO2 emissions. This study marks the first exploration of applying KD to the problem of individual vertebrae segmentation in CT, demonstrating the feasibility of achieving comparable performance to existing methods using smaller neural networks.

A Multifunctional Network with Uncertainty Estimation and Attention-Based Knowledge Distillation to Address Practical Challenges in Respiration Rate Estimation

Respiration rate is a vital parameter to indicate good health, wellbeing, and performance. As the estimation through classical measurement modes are limited only to rest or during slow movements, respiration rate is commonly estimated through physiological signals such as electrocardiogram and photoplethysmography due to the unobtrusive nature of wearable devices. Deep learning methodologies have gained much traction in the recent past to enhance accuracy during activities involving a lot of movement. However, these methods pose challenges, including model interpretability, uncertainty estimation in the context of respiration rate estimation, and model compactness in terms of deployment in wearable platforms. In this direction, we propose a multifunctional framework, which includes the combination of an attention mechanism, an uncertainty estimation functionality, and a knowledge distillation framework. We evaluated the performance of our framework on two datasets containing ambulatory movement. The attention mechanism visually and quantitatively improved instantaneous respiration rate estimation. Using Monte Carlo dropouts to embed the network with inferential uncertainty estimation resulted in the rejection of 3.7% of windows with high uncertainty, which consequently resulted in an overall reduction of 7.99% in the mean absolute error. The attention-aware knowledge distillation mechanism reduced the model's parameter count and inference time by 49.5% and 38.09%, respectively, without any increase in error rates. Through experimentation, ablation, and visualization, we demonstrated the efficacy of the proposed framework in addressing practical challenges, thus taking a step towards deployment in wearable edge devices.

Knowledge distillation with ensembles of convolutional neural networks for medical image segmentation

Purpose: Ensembles of convolutional neural networks (CNNs) often outperform a single CNN in medical image segmentation tasks, but inference is computationally more expensive and makes ensembles unattractive for some applications. We compared the performance of differently constructed ensembles with the performance of CNNs derived from these ensembles using knowledge distillation, a technique for reducing the footprint of large models such as ensembles. Approach: We investigated two different types of ensembles, namely, diverse ensembles of networks with three different architectures and two different loss-functions, and uniform ensembles of networks with the same architecture but initialized with different random seeds. For each ensemble, additionally, a single student network was trained to mimic the class probabilities predicted by the teacher model, the ensemble. We evaluated the performance of each network, the ensembles, and the corresponding distilled networks across three different publicly available datasets. These included chest computed tomography scans with four annotated organs of interest, brain magnetic resonance imaging (MRI) with six annotated brain structures, and cardiac cine-MRI with three annotated heart structures. Results: Both uniform and diverse ensembles obtained better results than any of the individual networks in the ensemble. Furthermore, applying knowledge distillation resulted in a single network that was smaller and faster without compromising performance compared with the ensemble it learned from. The distilled networks significantly outperformed the same network trained with reference segmentation instead of knowledge distillation. Conclusion: Knowledge distillation can compress segmentation ensembles of uniform or diverse composition into a single CNN while maintaining the performance of the ensemble.