Browsing by Subject "Deep Learning"

Purpose

Methods

Results

Conclusions

Glaucoma is a group of chronic eye diseases characterized by optic neuropathy, which causes irreversible vision loss. It is caused by progressive degeneration of the optic nerve, leading to gradual loss of the visual field from the periphery to the center, resulting in blindness if left untreated. Since the changes are gradual and the damage progresses generally slowly, glaucoma development is insidious and often diagnosed until it reaches an advanced stage. Early detection of glaucoma progression is necessary to monitor the atrophy and formulate treatment strategies to halt progressive functional vision impairments. The availability of data centric methods have made it possible for researchers to develop computer-aided algorithms for the clinical diagnosis of glaucoma and capture accurate disease characteristics. In this research, we use deep learning models, one such forefront, to identify complex disease characteristics and progression criteria, enabling the detection of subtle changes indicative of glaucoma progression.

To this end, we investigate the structure-function relationship of glaucoma progression and explore the possibility of predicting functional impairment from structural eye deterioration. We also analyze various statistical and machine-learning methods that have aided previous attempts to estimate progression, including emerging deep-learning techniques that use structural features like optical coherence tomography (OCT) scans to predict glaucoma progression accurately. We show through our investigations that these methods are still prone to confounding risk factors, especially variability due to age, data imbalances, potential noisy labels, lack of gold standard criteria, etc. We developed novel semi-supervised time-series algorithms to overcome these multifaceted challenges using unique data-driven approaches:

Weakly-Supervised Time-Series Learning: We develop a convolutional neural network-long short-term memory (CNN-LSTM) base model to encode the spatiotemporal features from the OCT scan sequence taken over a fixed follow-up. We model the rest of the deep learning architecture on the fact that original OCT sequences exhibit age-related progression, and reshuffling the sequence order, along with the knowledge of healthy eyes from a positive-unlabeled dataset, can establish robust pseudo-progression criteria for glaucoma. This circumvents the need for gold standard labels for disease progression.

Semi-supervised Time-Series Learning: We extend the above notion to a labeled case where labels are obtained from Guided Progression Analysis (GPA), a well-known, stable, and accurate functional assessment for glaucoma progression, but might be prone to noisy labels due to nuances in data acquisition. We model the age-related structural progression as a pseudo-identifier for glaucoma progression. We use this knowledge in a contrastive learning scheme where the foundational CNN-LSTM base learns accurate spatiotemporal characteristics from potentially mislabeled data and improves predictions.

Finally, we compare and show that these methods outperform conventional and state-of-the-art techniques.

Background

Purpose

Methods

Results

Conclusions

Purpose

Method

Results

Conclusion

Managing respiratory motion in radiotherapy for abdominal and lung stereotactic body radiation therapy (SBRT) patients is crucial to achieving dose conformity and sparing healthy tissue. While previous studies have utilized principal component analysis (PCA) combined with deep learning to localize lung tumors in real-time, these models lack testing or validation with patient data from later treatment fractions. This study aims to achieve highly accurate 3D tumor localization for abdominal and lung cancer patients using a PCA-based convolutional neural network (CNN) motion model. Another goal is to enhance this prediction model by addressing interfractional motion in lung patients through deep learning and multiple treatment-acquired CT datasets.

The patient's 4D computed tomography (4DCT) image was registered, and resulting deformation vector fields (DVFs) were approximated using PCA. PCA coefficients, linked to diaphragm displacement, controlled breath variability in synthetic CT images. Digitally reconstructed radiographs (DRRs) from synthetic CTs served as network input. The networks were evaluated using a CT image designated for abdominal and lung cancer patient testing. A DRR of the testing CT was input into the model, and the predicted CT image was validated against the test CT to locate the tumor.

This study validated a PCA-based CNN motion model for abdominal and lung cancer patients. Intrafractional motion modeling accurately predicted abdominal tumors with a maximum error of 1.41 mm, whereas lung tumor localization resulted in a maximum error of 2.83 mm. Interfractional motion significantly influenced the accuracy of lung tumor prediction.

Deep learning has enabled remarkable performance breakthroughs across various domains, including computer vision, natural language processing, and recommender systems. However, the typical deep neural network (DNN) models employed in these applications require millions of parameters and billions of operations, leading to substantial computational and memory requirements. While researchers have proposed compression methods, optimized frameworks, and specialized accelerators to improve efficiency, outstanding challenges persist, limiting the achievable gains.

A fundamental challenge lies in the inherent irregularity and sparsity of DNNs. Although these models exhibit significant sparsity, with a considerable fraction of weights and activations being zero or near-zero values, exploiting this sparsity efficiently on modern hardware is problematic due to the irregular distribution of non-zero elements. This irregularity leads to substantial overhead in indexing, gathering, and processing sparse data, resulting in poor utilization of computational and memory resources. Furthermore, recent research has identified a significant gap between the theoretical and practical improvements achieved by compression methods. Additionally, emerging DNN architectures with novel operators often nullify previous optimization efforts in software frameworks and hardware accelerators, necessitating continuous adaptation.

To address these critical challenges, this dissertation targets building a holistic approach that jointly optimizes algorithms, hardware architectures, and system designs to enable efficient deployment of DNNs in the presence of irregularity and sparsity. On the algorithm level, a novel hardware-friendly compression method based on matrix decomposition is proposed. The original convolutional kernels are decomposed into common basis kernels and a series of coefficients, with conventional pruning applied to the coefficients. This compressed DNN forms a hardware-friendly structure where the sparsity pattern is shared across input feature map pixels, alleviating sparse pattern processing costs.

On the hardware level, a novel sparse DNN accelerator is introduced to support the inference of the compressed DNN. Low-precision quantization is applied to sparse coefficients, and high-precision to basis kernels. By involving only low-precision coefficients in sparse processing, the hardware efficiently matches non-zero weights and activations using inverted butterfly networks. The shared basis kernels and sparse coefficients significantly reduce buffer size and bandwidth requirements, boosting performance and energy efficiency.

At the system level, a near-data processing framework is proposed to address the challenge of training large DNN-based recommendation models. This framework adopts computational storage devices and coherent system interconnects to partition the model into subtasks. Data-intensive embedding operations run on computational storage devices with customized memory hierarchies, while compute-intensive feature processing and aggregation operations are assigned to GPUs for maximum efficiency. This framework enables training large DNN-based recommendation models without expensive hardware investments.

Through joint optimization across algorithms, hardware architectures, and system designs, this research aims to overcome the limitations imposed by irregularity and sparsity, enabling efficient deployment of DNNs in a broad range of applications and resource-constrained environments. By addressing these critical issues, this work paves the way for fully harnessing the potential of deep learning technologies in practical settings.

Background

Methods

Results

Conclusions